sarvamai 1.1.16-alpha.3 → 1.1.17

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (48) hide show
  1. package/dist/cjs/BaseClient.js +2 -2
  2. package/dist/cjs/api/resources/chat/client/Client.js +8 -0
  3. package/dist/cjs/api/resources/chat/client/requests/CreateChatCompletionV2Request.d.ts +3 -3
  4. package/dist/cjs/api/resources/decisions/client/Client.d.ts +12 -4
  5. package/dist/cjs/api/resources/decisions/client/Client.js +12 -4
  6. package/dist/cjs/api/resources/responses/client/Client.js +4 -0
  7. package/dist/cjs/api/resources/responses/client/requests/ResponsesCreateRequest.d.ts +5 -1
  8. package/dist/cjs/api/resources/responses/client/requests/ResponsesCreateStreamRequest.d.ts +5 -1
  9. package/dist/cjs/api/types/Choice.d.ts +1 -1
  10. package/dist/cjs/api/types/CreateResponseV2Request.d.ts +5 -1
  11. package/dist/cjs/api/types/ModelObject.d.ts +10 -0
  12. package/dist/cjs/api/types/ModelObject.js +9 -0
  13. package/dist/cjs/api/types/ReasoningEffort.d.ts +2 -1
  14. package/dist/cjs/api/types/ReasoningEffort.js +2 -1
  15. package/dist/cjs/api/types/ResponseReasoning.d.ts +1 -1
  16. package/dist/cjs/api/types/ResponseReasoningEffort.d.ts +2 -1
  17. package/dist/cjs/api/types/ResponseReasoningEffort.js +2 -1
  18. package/dist/cjs/api/types/ResponsesModelIds.d.ts +7 -1
  19. package/dist/cjs/api/types/ResponsesModelIds.js +7 -1
  20. package/dist/cjs/api/types/V2ModelIds.d.ts +8 -2
  21. package/dist/cjs/api/types/V2ModelIds.js +8 -2
  22. package/dist/cjs/version.d.ts +1 -1
  23. package/dist/cjs/version.js +1 -1
  24. package/dist/esm/BaseClient.mjs +2 -2
  25. package/dist/esm/api/resources/chat/client/Client.mjs +8 -0
  26. package/dist/esm/api/resources/chat/client/requests/CreateChatCompletionV2Request.d.mts +3 -3
  27. package/dist/esm/api/resources/decisions/client/Client.d.mts +12 -4
  28. package/dist/esm/api/resources/decisions/client/Client.mjs +12 -4
  29. package/dist/esm/api/resources/responses/client/Client.mjs +4 -0
  30. package/dist/esm/api/resources/responses/client/requests/ResponsesCreateRequest.d.mts +5 -1
  31. package/dist/esm/api/resources/responses/client/requests/ResponsesCreateStreamRequest.d.mts +5 -1
  32. package/dist/esm/api/types/Choice.d.mts +1 -1
  33. package/dist/esm/api/types/CreateResponseV2Request.d.mts +5 -1
  34. package/dist/esm/api/types/ModelObject.d.mts +10 -0
  35. package/dist/esm/api/types/ModelObject.mjs +8 -1
  36. package/dist/esm/api/types/ReasoningEffort.d.mts +2 -1
  37. package/dist/esm/api/types/ReasoningEffort.mjs +2 -1
  38. package/dist/esm/api/types/ResponseReasoning.d.mts +1 -1
  39. package/dist/esm/api/types/ResponseReasoningEffort.d.mts +2 -1
  40. package/dist/esm/api/types/ResponseReasoningEffort.mjs +2 -1
  41. package/dist/esm/api/types/ResponsesModelIds.d.mts +7 -1
  42. package/dist/esm/api/types/ResponsesModelIds.mjs +7 -1
  43. package/dist/esm/api/types/V2ModelIds.d.mts +8 -2
  44. package/dist/esm/api/types/V2ModelIds.mjs +8 -2
  45. package/dist/esm/version.d.mts +1 -1
  46. package/dist/esm/version.mjs +1 -1
  47. package/package.json +1 -1
  48. package/reference.md +14 -6
@@ -43,8 +43,8 @@ function normalizeClientOptions(options) {
43
43
  const headers = (0, headers_js_1.mergeHeaders)({
44
44
  "X-Fern-Language": "JavaScript",
45
45
  "X-Fern-SDK-Name": "sarvamai",
46
- "X-Fern-SDK-Version": "1.1.16-alpha.3",
47
- "User-Agent": "sarvamai/1.1.16-alpha.3",
46
+ "X-Fern-SDK-Version": "1.1.17",
47
+ "User-Agent": "sarvamai/1.1.17",
48
48
  "X-Fern-Runtime": core.RUNTIME.type,
49
49
  "X-Fern-Runtime-Version": core.RUNTIME.version,
50
50
  }, options === null || options === void 0 ? void 0 : options.headers);
@@ -341,8 +341,12 @@ class ChatClient {
341
341
  throw new SarvamAI.BadRequestError(body, rawResponse);
342
342
  case 401:
343
343
  throw new SarvamAI.UnauthorizedError(body, rawResponse);
344
+ case 402:
345
+ throw new SarvamAI.PaymentRequiredError(body, rawResponse);
344
346
  case 403:
345
347
  throw new SarvamAI.ForbiddenError(body, rawResponse);
348
+ case 404:
349
+ throw new SarvamAI.NotFoundError(body, rawResponse);
346
350
  case 413:
347
351
  throw new SarvamAI.ContentTooLargeError(body, rawResponse);
348
352
  case 422:
@@ -351,8 +355,12 @@ class ChatClient {
351
355
  throw new SarvamAI.TooManyRequestsError(body, rawResponse);
352
356
  case 500:
353
357
  throw new SarvamAI.InternalServerError(body, rawResponse);
358
+ case 502:
359
+ throw new SarvamAI.BadGatewayError(body, rawResponse);
354
360
  case 503:
355
361
  throw new SarvamAI.ServiceUnavailableError(body, rawResponse);
362
+ case 504:
363
+ throw new SarvamAI.GatewayTimeoutError(body, rawResponse);
356
364
  default:
357
365
  throw new errors.SarvamAIError({
358
366
  statusCode,
@@ -20,7 +20,7 @@ export interface CreateChatCompletionV2Request {
20
20
  temperature?: number;
21
21
  /** Nucleus sampling: greater than `0` and at most `1`. On `glm5.3`, `top_p: 0` is rejected. Prefer changing either `top_p` or `temperature`, not both. */
22
22
  top_p?: number;
23
- /** On reasoning models (`glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`): **`low`**, **`high`**, or **`max`**. Default when omitted is **`max`**, except **`high`** on `glm5.3-flash`, `deepseekv4.1-flash-1017`, and `deepseekv4-pro`, and no reasoning on `gemma4`. **`none`** turns thinking off on `deepseekv4.1-flash-1017` and `deepseekv4-pro`. `kimi-k3.1-pro` always thinks. Trace in `message.reasoning_content`; billed as completion tokens and counts toward `max_tokens`. On `gemma4`, reasoning is off unless this is set (`low`, `high`, or `max`). */
23
+ /** On reasoning models: **`low`**, **`high`**, or **`max`** on every reasoning model, and **`medium`** on `sarvam-105b`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, `kimi-k3.1-pro`, and `qwen38-27b` (`kimi-k3.1-pro` runs `medium` as `high`). Default when omitted is **`max`** on `glm5.3` and `kimi-k3.1-pro`, **`high`** on `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro`, and no reasoning on `gemma4`; `qwen38-27b` thinks by default. **`none`** turns thinking off on `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`. `kimi-k3.1-pro` always thinks: `none`, `minimal`, and `xhigh` return `400`. Trace in `message.reasoning_content`; billed as completion tokens and counts toward `max_tokens`. On `gemma4`, reasoning is off unless this is set (`low`, `high`, or `max`). */
24
24
  reasoning_effort?: SarvamAI.ReasoningEffort;
25
25
  /** Maximum tokens to generate in the completion, including reasoning tokens on reasoning models. Default `2048`. Set explicitly when you need longer answers — a low limit can yield `finish_reason: "length"` with little or no visible `content` if reasoning consumes the budget. */
26
26
  max_tokens?: number;
@@ -59,8 +59,8 @@ export interface CreateChatCompletionV2Request {
59
59
  parallel_tool_calls?: boolean;
60
60
  /** Only valid with `stream: true` (otherwise `400`). Must be a top-level field, not inside `extra_body`. Use `{"include_usage": true}` to receive token usage in the final SSE chunk. */
61
61
  stream_options?: SarvamAI.ChatStreamOptions;
62
- /** If `true`, return log probabilities for output tokens (use with `top_logprobs`). Supported on **`sarvam-105b`**, **`gemma4`**, **`glm5.3-flash`**, **`deepseekv4.1-flash-1017`**, **`deepseekv4-pro`**, and **`kimi-k3.1-pro`**. Not supported on **`glm5.3`** (`400` if enabled). */
62
+ /** If `true`, return log probabilities for output tokens (use with `top_logprobs`). Supported on **`sarvam-105b`**, **`gemma4`**, **`glm5.3-flash`**, **`deepseekv4.1-flash-1017`**, **`deepseekv4.1-flash-1017-interactive`**, **`deepseekv4-pro`**, and **`qwen38-27b`**. Not supported on **`glm5.3`** or **`kimi-k3.1-pro`** (`400` if enabled). */
63
63
  logprobs?: boolean;
64
- /** How many top token log probabilities to return at each position. Takes effect only when `logprobs` is `true`; without it, some models ignore the field and others return **`400`**. Same model support as `logprobs`: works on `sarvam-105b`, `gemma4`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro`; on `glm5.3`, omit entirely — do not send `0`, since many SDKs add it by default. */
64
+ /** How many top token log probabilities to return at each position. Takes effect only when `logprobs` is `true`; without it, some models ignore the field and others return **`400`**. Same model support as `logprobs`: works on `sarvam-105b`, `gemma4`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`; on `glm5.3` and `kimi-k3.1-pro`, omit entirely — do not send `0`, since many SDKs add it by default. */
65
65
  top_logprobs?: number;
66
66
  }
@@ -11,17 +11,25 @@ export declare class DecisionsClient {
11
11
  protected readonly _options: NormalizedClientOptionsWithAuth<DecisionsClient.Options>;
12
12
  constructor(options?: DecisionsClient.Options);
13
13
  /**
14
- * Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse. The order of `choice` options does not affect the answer.
14
+ * Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse.
15
+ *
16
+ * **Models:** `sarvam-prism` (serverless) and `clef` (Provisioned Throughput; also reads images).
15
17
  *
16
18
  * Requires `api-subscription-key`, or `Authorization: Bearer <key>`.
17
19
  *
18
20
  * **Question types:** `noul` (yes/no), `choice` (one of your options) and `score` (a level on your ordered scale).
19
21
  *
20
- * **Repeatability:** probabilities can vary slightly between identical requests; the top answer is stable.
22
+ * **Repeatability:** on `sarvam-prism`, probabilities can vary slightly between identical requests and the top answer is stable; the order of `choice` options does not affect the answer. On `clef`, identical requests return identical probabilities.
23
+ *
24
+ * **Images:** `clef` reads base64 images in `images`; image URLs return `400`. `sarvam-prism` refuses `images` (image URLs return `400`, inline images `422`).
25
+ *
26
+ * **Not supported:** streaming. `temperature` returns `422` on both models. `sarvam-prism` also refuses `max_tokens` and `stream` with `422`; `clef` accepts and ignores them. Unknown keys inside a question return `422`.
21
27
  *
22
- * **Not supported:** streaming, and images on `sarvam-prism` (`clef` accepts base64 images). Generation fields such as `temperature`, `max_tokens` and `stream` are refused with `422`. On `sarvam-prism`, `images` is refused too (image URLs return `400`, inline images `422`); `clef` reads base64 images and refuses image URLs with `400`. Unknown keys inside a question also return `422`.
28
+ * **Limits:**
29
+ * - `sarvam-prism`: 32,768 tokens per question (state + question + options), 2–255 options per `choice`, and 2–10 levels per `score`. If any one question breaks a limit, the whole request fails with `422`.
30
+ * - `clef`: 65,536 tokens per request (state, every question and its options, and images), 1–255 options per `choice`, and 1–10 levels per `score`. A longer request fails with `400`.
23
31
  *
24
- * **Limits:** 32,768 tokens per question (state + question + options), 2–255 options per `choice` and 2–10 levels per `score`. There is no cap on the number of questions. The state is not truncated, and if any one question breaks a limit the whole request fails with `422`.
32
+ * There is no cap on the number of questions, and the state is never truncated.
25
33
  *
26
34
  * **Timeout:** each request must finish within 60 seconds. A longer request fails with `503 model_overloaded`; split very large batches across several requests before retrying.
27
35
  *
@@ -56,17 +56,25 @@ class DecisionsClient {
56
56
  this._options = (0, BaseClient_js_1.normalizeClientOptionsWithAuth)(options);
57
57
  }
58
58
  /**
59
- * Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse. The order of `choice` options does not affect the answer.
59
+ * Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse.
60
+ *
61
+ * **Models:** `sarvam-prism` (serverless) and `clef` (Provisioned Throughput; also reads images).
60
62
  *
61
63
  * Requires `api-subscription-key`, or `Authorization: Bearer <key>`.
62
64
  *
63
65
  * **Question types:** `noul` (yes/no), `choice` (one of your options) and `score` (a level on your ordered scale).
64
66
  *
65
- * **Repeatability:** probabilities can vary slightly between identical requests; the top answer is stable.
67
+ * **Repeatability:** on `sarvam-prism`, probabilities can vary slightly between identical requests and the top answer is stable; the order of `choice` options does not affect the answer. On `clef`, identical requests return identical probabilities.
68
+ *
69
+ * **Images:** `clef` reads base64 images in `images`; image URLs return `400`. `sarvam-prism` refuses `images` (image URLs return `400`, inline images `422`).
70
+ *
71
+ * **Not supported:** streaming. `temperature` returns `422` on both models. `sarvam-prism` also refuses `max_tokens` and `stream` with `422`; `clef` accepts and ignores them. Unknown keys inside a question return `422`.
66
72
  *
67
- * **Not supported:** streaming, and images on `sarvam-prism` (`clef` accepts base64 images). Generation fields such as `temperature`, `max_tokens` and `stream` are refused with `422`. On `sarvam-prism`, `images` is refused too (image URLs return `400`, inline images `422`); `clef` reads base64 images and refuses image URLs with `400`. Unknown keys inside a question also return `422`.
73
+ * **Limits:**
74
+ * - `sarvam-prism`: 32,768 tokens per question (state + question + options), 2–255 options per `choice`, and 2–10 levels per `score`. If any one question breaks a limit, the whole request fails with `422`.
75
+ * - `clef`: 65,536 tokens per request (state, every question and its options, and images), 1–255 options per `choice`, and 1–10 levels per `score`. A longer request fails with `400`.
68
76
  *
69
- * **Limits:** 32,768 tokens per question (state + question + options), 2–255 options per `choice` and 2–10 levels per `score`. There is no cap on the number of questions. The state is not truncated, and if any one question breaks a limit the whole request fails with `422`.
77
+ * There is no cap on the number of questions, and the state is never truncated.
70
78
  *
71
79
  * **Timeout:** each request must finish within 60 seconds. A longer request fails with `503 model_overloaded`; split very large batches across several requests before retrying.
72
80
  *
@@ -175,6 +175,8 @@ class ResponsesClient {
175
175
  throw new SarvamAI.BadRequestError(body, rawResponse);
176
176
  case 401:
177
177
  throw new SarvamAI.UnauthorizedError(body, rawResponse);
178
+ case 402:
179
+ throw new SarvamAI.PaymentRequiredError(body, rawResponse);
178
180
  case 403:
179
181
  throw new SarvamAI.ForbiddenError(body, rawResponse);
180
182
  case 404:
@@ -221,6 +223,8 @@ class ResponsesClient {
221
223
  throw new SarvamAI.BadRequestError(_response.error.body, _response.rawResponse);
222
224
  case 401:
223
225
  throw new SarvamAI.UnauthorizedError(_response.error.body, _response.rawResponse);
226
+ case 402:
227
+ throw new SarvamAI.PaymentRequiredError(_response.error.body, _response.rawResponse);
224
228
  case 403:
225
229
  throw new SarvamAI.ForbiddenError(_response.error.body, _response.rawResponse);
226
230
  case 404:
@@ -19,7 +19,7 @@ export interface ResponsesCreateRequest {
19
19
  tools?: SarvamAI.ResponseFunctionTool[];
20
20
  /** Forwarded exactly as written: `none`, `auto`, `required`, `{ "type": "function", "name": "..." }`, or another object. */
21
21
  tool_choice?: ResponsesCreateRequest.ToolChoice;
22
- /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro`, `low`, `high`, or `max` all work directly. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
22
+ /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, the DeepSeek models, `kimi-k3.1-pro`, and `qwen38-27b`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro`, `low`, `high`, `max`, or `none` (thinking off); on `kimi-k3.1-pro`, `low`, `medium`, `high`, or `max` (thinking is always on); on `qwen38-27b`, `none` turns thinking off. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
23
23
  reasoning?: SarvamAI.ResponseReasoning;
24
24
  /** Merged into the outbound body. Must not overwrite any modelled field (`400`). `max_tool_calls` and `prompt` are `400` here and in `extra_body` — not supported by any serving deployment. */
25
25
  extra_body?: Record<string, unknown>;
@@ -48,6 +48,10 @@ export interface ResponsesCreateRequest {
48
48
  background?: boolean;
49
49
  /** `400 invalid_request_error` — not supported by any serving deployment, top-level or via `extra_body`. */
50
50
  max_tool_calls?: number;
51
+ /** How many of the most likely tokens to return at each output position, with their log probabilities. Send it with `include: ["message.output_text.logprobs"]`; without `include`, `output[].content[].logprobs` comes back empty. `0` returns no logprobs. Supported on `sarvam-105b`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`; `glm5.3`, `gemma4`, and `kimi-k3.1-pro` return `400`. */
52
+ top_logprobs?: number;
53
+ /** Extra data to add to the response. `message.output_text.logprobs` returns token log probabilities (use with `top_logprobs`); `reasoning.encrypted_content` is accepted. An unknown value returns `400`. */
54
+ include?: string[];
51
55
  /** Accepts any additional properties */
52
56
  [key: string]: any;
53
57
  }
@@ -19,7 +19,7 @@ export interface ResponsesCreateStreamRequest {
19
19
  tools?: SarvamAI.ResponseFunctionTool[];
20
20
  /** Forwarded exactly as written: `none`, `auto`, `required`, `{ "type": "function", "name": "..." }`, or another object. */
21
21
  tool_choice?: ResponsesCreateStreamRequest.ToolChoice;
22
- /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro`, `low`, `high`, or `max` all work directly. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
22
+ /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, the DeepSeek models, `kimi-k3.1-pro`, and `qwen38-27b`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro`, `low`, `high`, `max`, or `none` (thinking off); on `kimi-k3.1-pro`, `low`, `medium`, `high`, or `max` (thinking is always on); on `qwen38-27b`, `none` turns thinking off. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
23
23
  reasoning?: SarvamAI.ResponseReasoning;
24
24
  /** Merged into the outbound body. Must not overwrite any modelled field (`400`). `max_tool_calls` and `prompt` are `400` here and in `extra_body` — not supported by any serving deployment. */
25
25
  extra_body?: Record<string, unknown>;
@@ -48,6 +48,10 @@ export interface ResponsesCreateStreamRequest {
48
48
  background?: boolean;
49
49
  /** `400 invalid_request_error` — not supported by any serving deployment, top-level or via `extra_body`. */
50
50
  max_tool_calls?: number;
51
+ /** How many of the most likely tokens to return at each output position, with their log probabilities. Send it with `include: ["message.output_text.logprobs"]`; without `include`, `output[].content[].logprobs` comes back empty. `0` returns no logprobs. Supported on `sarvam-105b`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`; `glm5.3`, `gemma4`, and `kimi-k3.1-pro` return `400`. */
52
+ top_logprobs?: number;
53
+ /** Extra data to add to the response. `message.output_text.logprobs` returns token log probabilities (use with `top_logprobs`); `reasoning.encrypted_content` is accepted. An unknown value returns `400`. */
54
+ include?: string[];
51
55
  /** Accepts any additional properties */
52
56
  [key: string]: any;
53
57
  }
@@ -9,7 +9,7 @@ export interface Choice {
9
9
  finish_reason: SarvamAI.FinishReason;
10
10
  /** The index of the choice in the list of choices. */
11
11
  index: number;
12
- /** Token log probabilities when supported (`sarvam-105b`, `gemma4`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`). Always `null` on `glm5.3`, which doesn't support this field. */
12
+ /** Token log probabilities when supported (`sarvam-105b`, `gemma4`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, `qwen38-27b`). Not supported on `glm5.3` or `kimi-k3.1-pro`, which return `400` when `logprobs` is set. */
13
13
  logprobs?: Record<string, unknown> | undefined;
14
14
  message: SarvamAI.ChatCompletionResponseMessage;
15
15
  }
@@ -14,7 +14,7 @@ export interface CreateResponseV2Request {
14
14
  tools?: SarvamAI.ResponseFunctionTool[] | undefined;
15
15
  /** Forwarded exactly as written: `none`, `auto`, `required`, `{ "type": "function", "name": "..." }`, or another object. */
16
16
  tool_choice?: CreateResponseV2Request.ToolChoice | undefined;
17
- /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro`, `low`, `high`, or `max` all work directly. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
17
+ /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, the DeepSeek models, `kimi-k3.1-pro`, and `qwen38-27b`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro`, `low`, `high`, `max`, or `none` (thinking off); on `kimi-k3.1-pro`, `low`, `medium`, `high`, or `max` (thinking is always on); on `qwen38-27b`, `none` turns thinking off. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
18
18
  reasoning?: SarvamAI.ResponseReasoning | undefined;
19
19
  /** Merged into the outbound body. Must not overwrite any modelled field (`400`). `max_tool_calls` and `prompt` are `400` here and in `extra_body` — not supported by any serving deployment. */
20
20
  extra_body?: Record<string, unknown> | undefined;
@@ -43,6 +43,10 @@ export interface CreateResponseV2Request {
43
43
  background?: boolean | undefined;
44
44
  /** `400 invalid_request_error` — not supported by any serving deployment, top-level or via `extra_body`. */
45
45
  max_tool_calls?: number | undefined;
46
+ /** How many of the most likely tokens to return at each output position, with their log probabilities. Send it with `include: ["message.output_text.logprobs"]`; without `include`, `output[].content[].logprobs` comes back empty. `0` returns no logprobs. Supported on `sarvam-105b`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`; `glm5.3`, `gemma4`, and `kimi-k3.1-pro` return `400`. */
47
+ top_logprobs?: number | undefined;
48
+ /** Extra data to add to the response. `message.output_text.logprobs` returns token log probabilities (use with `top_logprobs`); `reasoning.encrypted_content` is accepted. An unknown value returns `400`. */
49
+ include?: string[] | undefined;
46
50
  /** Accepts any additional properties */
47
51
  [key: string]: any;
48
52
  }
@@ -3,4 +3,14 @@ export interface ModelObject {
3
3
  object?: "model" | undefined;
4
4
  created: number;
5
5
  owned_by: string;
6
+ /** How the model is offered. `"provisioned"`: Provisioned Throughput, enabled per organisation (without access, `deepseekv4-pro` returns `400` and the others return `404` and aren't listed). `"standard"`: serverless, any API key can call it. */
7
+ access?: ModelObject.Access | undefined;
8
+ }
9
+ export declare namespace ModelObject {
10
+ /** How the model is offered. `"provisioned"`: Provisioned Throughput, enabled per organisation (without access, `deepseekv4-pro` returns `400` and the others return `404` and aren't listed). `"standard"`: serverless, any API key can call it. */
11
+ const Access: {
12
+ readonly Standard: "standard";
13
+ readonly Provisioned: "provisioned";
14
+ };
15
+ type Access = (typeof Access)[keyof typeof Access];
6
16
  }
@@ -1,3 +1,12 @@
1
1
  "use strict";
2
2
  // This file was auto-generated by Fern from our API Definition.
3
3
  Object.defineProperty(exports, "__esModule", { value: true });
4
+ exports.ModelObject = void 0;
5
+ var ModelObject;
6
+ (function (ModelObject) {
7
+ /** How the model is offered. `"provisioned"`: Provisioned Throughput, enabled per organisation (without access, `deepseekv4-pro` returns `400` and the others return `404` and aren't listed). `"standard"`: serverless, any API key can call it. */
8
+ ModelObject.Access = {
9
+ Standard: "standard",
10
+ Provisioned: "provisioned",
11
+ };
12
+ })(ModelObject || (exports.ModelObject = ModelObject = {}));
@@ -1,7 +1,8 @@
1
- /** Reasoning depth on `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`: `low`, `high`, or `max`. Default when omitted is `max` (`high` on `glm5.3-flash`, `deepseekv4.1-flash-1017`, and `deepseekv4-pro`; no reasoning on `gemma4`). `none` turns thinking off on `deepseekv4.1-flash-1017` and `deepseekv4-pro`; other models reject it or keep thinking. */
1
+ /** Reasoning depth. `low`, `high`, and `max` work on every reasoning model; `medium` also works on `sarvam-105b`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, `kimi-k3.1-pro`, and `qwen38-27b`. Default when omitted is `max` on `glm5.3` and `kimi-k3.1-pro`, `high` on `glm5.3-flash` and the DeepSeek models, and no reasoning on `gemma4`. `none` turns thinking off on the DeepSeek models and `qwen38-27b`; `kimi-k3.1-pro` refuses it, and other models reject it or keep thinking. */
2
2
  export declare const ReasoningEffort: {
3
3
  readonly None: "none";
4
4
  readonly Low: "low";
5
+ readonly Medium: "medium";
5
6
  readonly High: "high";
6
7
  readonly Max: "max";
7
8
  };
@@ -2,10 +2,11 @@
2
2
  // This file was auto-generated by Fern from our API Definition.
3
3
  Object.defineProperty(exports, "__esModule", { value: true });
4
4
  exports.ReasoningEffort = void 0;
5
- /** Reasoning depth on `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`: `low`, `high`, or `max`. Default when omitted is `max` (`high` on `glm5.3-flash`, `deepseekv4.1-flash-1017`, and `deepseekv4-pro`; no reasoning on `gemma4`). `none` turns thinking off on `deepseekv4.1-flash-1017` and `deepseekv4-pro`; other models reject it or keep thinking. */
5
+ /** Reasoning depth. `low`, `high`, and `max` work on every reasoning model; `medium` also works on `sarvam-105b`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, `kimi-k3.1-pro`, and `qwen38-27b`. Default when omitted is `max` on `glm5.3` and `kimi-k3.1-pro`, `high` on `glm5.3-flash` and the DeepSeek models, and no reasoning on `gemma4`. `none` turns thinking off on the DeepSeek models and `qwen38-27b`; `kimi-k3.1-pro` refuses it, and other models reject it or keep thinking. */
6
6
  exports.ReasoningEffort = {
7
7
  None: "none",
8
8
  Low: "low",
9
+ Medium: "medium",
9
10
  High: "high",
10
11
  Max: "max",
11
12
  };
@@ -1,6 +1,6 @@
1
1
  import type * as SarvamAI from "../index.js";
2
2
  export interface ResponseReasoning {
3
- /** `low`, `high`, or `max`. On **`glm5.3`** and **`glm5.3-flash`**, sending the literal string `max` returns `400`; omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. **`deepseekv4.1-flash-1017`**, **`deepseekv4-pro`**, and **`kimi-k3.1-pro`** accept `max` directly on this route. `none` turns thinking off on `deepseekv4-pro`. `gemma4` accepts `low` or `high`; omitting it means no reasoning. */
3
+ /** `low`, `high`, or `max`, plus `medium` and `none` on some models. On **`glm5.3`** and **`glm5.3-flash`**, sending the literal string `max` returns `400`; omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. **`deepseekv4.1-flash-1017`**, **`deepseekv4.1-flash-1017-interactive`**, and **`deepseekv4-pro`** accept `max` directly, and `none` turns thinking off. **`kimi-k3.1-pro`** accepts `low`, `medium`, `high`, or `max` and always thinks. **`qwen38-27b`** thinks by default; `none` turns it off. `gemma4` accepts `low` or `high`; omitting it means no reasoning. */
4
4
  effort?: SarvamAI.ResponseReasoningEffort | undefined;
5
5
  /** `auto` or `detailed`. Without it, the reasoning item's `summary` is empty; read the trace from `content[0].text`. */
6
6
  summary?: ResponseReasoning.Summary | undefined;
@@ -1,7 +1,8 @@
1
- /** Reasoning effort on the Responses route. `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro` accept all three values directly. `glm5.3` and `glm5.3-flash` accept `low` and `high`; the literal string `max` returns `400` on them. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. `none` turns thinking off on `deepseekv4-pro`. `gemma4` accepts `low` or `high`; omitting the field means no reasoning. */
1
+ /** Reasoning effort on the Responses route. `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro` accept `low`, `high`, and `max` directly, and `none` turns thinking off. `kimi-k3.1-pro` accepts `low`, `medium`, `high`, and `max` and always thinks. `qwen38-27b` thinks by default; `none` turns it off. `glm5.3` and `glm5.3-flash` accept `low` and `high`; the literal string `max` returns `400` on them. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. `gemma4` accepts `low` or `high`; omitting the field means no reasoning. */
2
2
  export declare const ResponseReasoningEffort: {
3
3
  readonly None: "none";
4
4
  readonly Low: "low";
5
+ readonly Medium: "medium";
5
6
  readonly High: "high";
6
7
  readonly Max: "max";
7
8
  };
@@ -2,10 +2,11 @@
2
2
  // This file was auto-generated by Fern from our API Definition.
3
3
  Object.defineProperty(exports, "__esModule", { value: true });
4
4
  exports.ResponseReasoningEffort = void 0;
5
- /** Reasoning effort on the Responses route. `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro` accept all three values directly. `glm5.3` and `glm5.3-flash` accept `low` and `high`; the literal string `max` returns `400` on them. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. `none` turns thinking off on `deepseekv4-pro`. `gemma4` accepts `low` or `high`; omitting the field means no reasoning. */
5
+ /** Reasoning effort on the Responses route. `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro` accept `low`, `high`, and `max` directly, and `none` turns thinking off. `kimi-k3.1-pro` accepts `low`, `medium`, `high`, and `max` and always thinks. `qwen38-27b` thinks by default; `none` turns it off. `glm5.3` and `glm5.3-flash` accept `low` and `high`; the literal string `max` returns `400` on them. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. `gemma4` accepts `low` or `high`; omitting the field means no reasoning. */
6
6
  exports.ResponseReasoningEffort = {
7
7
  None: "none",
8
8
  Low: "low",
9
+ Medium: "medium",
9
10
  High: "high",
10
11
  Max: "max",
11
12
  };
@@ -2,20 +2,26 @@
2
2
  * Model IDs accepted by `POST /v2/responses` (case-sensitive; unknown ids return `404`):
3
3
  *
4
4
  * - **`sarvam-105b`** — Sarvam flagship model.
5
+ * - **`sarvam-105b-conversations`** — Sarvam 105B tuned for real-time dialogue.
5
6
  * - **`glm5.3`** — 1,048,576-token context; reasoning in `output` as a `reasoning` item; flat function tools; max request size 10,485,760 bytes.
6
7
  * - **`glm5.3-flash`** — long context, vision (inline base64 images), tools, and reasoning; `reasoning.effort` accepts `low` or `high` only, as on `glm5.3`.
7
8
  * - **`gemma4`** — 262,144-token context, vision (inline base64 images), tools, and optional reasoning (`reasoning.effort` `low` or `high`).
8
- * - **`deepseekv4.1-flash-1017`** — long context, tools, and reasoning.
9
+ * - **`deepseekv4.1-flash-1017`** — long context, vision (inline base64 images), tools, and reasoning.
10
+ * - **`deepseekv4.1-flash-1017-interactive`** — same capabilities as `deepseekv4.1-flash-1017`, on a high-throughput deployment (more than 200 output tokens per second).
9
11
  * - **`deepseekv4-pro`** (Provisioned Throughput) — long context, text input, tools, and reasoning (default `high`).
10
12
  * - **`kimi-k3.1-pro`** (Provisioned Throughput) — long context, image input (inline base64; no video on this route), tools, and reasoning (always on).
13
+ * - **`qwen38-27b`** (Provisioned Throughput) — 262,144-token context, image input (inline base64), tools, and reasoning (on by default).
11
14
  */
12
15
  export declare const ResponsesModelIds: {
13
16
  readonly Sarvam105B: "sarvam-105b";
17
+ readonly Sarvam105BConversations: "sarvam-105b-conversations";
14
18
  readonly Glm53: "glm5.3";
15
19
  readonly Glm53Flash: "glm5.3-flash";
16
20
  readonly Gemma4: "gemma4";
17
21
  readonly Deepseekv41Flash1017: "deepseekv4.1-flash-1017";
22
+ readonly Deepseekv41Flash1017Interactive: "deepseekv4.1-flash-1017-interactive";
18
23
  readonly Deepseekv4Pro: "deepseekv4-pro";
19
24
  readonly KimiK31Pro: "kimi-k3.1-pro";
25
+ readonly Qwen3827B: "qwen38-27b";
20
26
  };
21
27
  export type ResponsesModelIds = (typeof ResponsesModelIds)[keyof typeof ResponsesModelIds];
@@ -6,19 +6,25 @@ exports.ResponsesModelIds = void 0;
6
6
  * Model IDs accepted by `POST /v2/responses` (case-sensitive; unknown ids return `404`):
7
7
  *
8
8
  * - **`sarvam-105b`** — Sarvam flagship model.
9
+ * - **`sarvam-105b-conversations`** — Sarvam 105B tuned for real-time dialogue.
9
10
  * - **`glm5.3`** — 1,048,576-token context; reasoning in `output` as a `reasoning` item; flat function tools; max request size 10,485,760 bytes.
10
11
  * - **`glm5.3-flash`** — long context, vision (inline base64 images), tools, and reasoning; `reasoning.effort` accepts `low` or `high` only, as on `glm5.3`.
11
12
  * - **`gemma4`** — 262,144-token context, vision (inline base64 images), tools, and optional reasoning (`reasoning.effort` `low` or `high`).
12
- * - **`deepseekv4.1-flash-1017`** — long context, tools, and reasoning.
13
+ * - **`deepseekv4.1-flash-1017`** — long context, vision (inline base64 images), tools, and reasoning.
14
+ * - **`deepseekv4.1-flash-1017-interactive`** — same capabilities as `deepseekv4.1-flash-1017`, on a high-throughput deployment (more than 200 output tokens per second).
13
15
  * - **`deepseekv4-pro`** (Provisioned Throughput) — long context, text input, tools, and reasoning (default `high`).
14
16
  * - **`kimi-k3.1-pro`** (Provisioned Throughput) — long context, image input (inline base64; no video on this route), tools, and reasoning (always on).
17
+ * - **`qwen38-27b`** (Provisioned Throughput) — 262,144-token context, image input (inline base64), tools, and reasoning (on by default).
15
18
  */
16
19
  exports.ResponsesModelIds = {
17
20
  Sarvam105B: "sarvam-105b",
21
+ Sarvam105BConversations: "sarvam-105b-conversations",
18
22
  Glm53: "glm5.3",
19
23
  Glm53Flash: "glm5.3-flash",
20
24
  Gemma4: "gemma4",
21
25
  Deepseekv41Flash1017: "deepseekv4.1-flash-1017",
26
+ Deepseekv41Flash1017Interactive: "deepseekv4.1-flash-1017-interactive",
22
27
  Deepseekv4Pro: "deepseekv4-pro",
23
28
  KimiK31Pro: "kimi-k3.1-pro",
29
+ Qwen3827B: "qwen38-27b",
24
30
  };
@@ -2,20 +2,26 @@
2
2
  * Model IDs accepted by `POST /v2/chat/completions` (case-sensitive; unknown ids return `404`):
3
3
  *
4
4
  * - **`sarvam-105b`** — Sarvam flagship chat model.
5
+ * - **`sarvam-105b-conversations`** — Sarvam 105B tuned for real-time dialogue and voice agents; 32,000-token context window.
5
6
  * - **`glm5.3`** — 1,048,576-token context window; visible reasoning in `reasoning_content` (billed as completion tokens); tool calling; max request size 10,485,760 bytes; `n` must be `1`.
6
7
  * - **`glm5.3-flash`** — 1,048,576-token context window, image input (inline base64), tool calling, and reasoning (on by default).
7
8
  * - **`gemma4`** — 262,144-token context window, image input (inline base64), tool calling, and optional reasoning (`reasoning_effort`).
8
- * - **`deepseekv4.1-flash-1017`** — 1,048,576-token context window, tool calling, and reasoning.
9
+ * - **`deepseekv4.1-flash-1017`** — 1,048,576-token context window, image input (inline base64), tool calling, and reasoning.
10
+ * - **`deepseekv4.1-flash-1017-interactive`** — the same model on a high-throughput deployment: more than 200 output tokens per second, against about 80; same capabilities as `deepseekv4.1-flash-1017`.
9
11
  * - **`deepseekv4-pro`** (Provisioned Throughput) — 1,000,000-token context window; text input only; tool calling; reasoning (default `high`; `none` turns it off).
10
- * - **`kimi-k3.1-pro`** (Provisioned Throughput) — 1,048,576-token context window; image input and short video clips (inline base64); tool calling; reasoning is always on (default `max`).
12
+ * - **`kimi-k3.1-pro`** (Provisioned Throughput) — 1,048,576-token context window; image and video input (inline base64); tool calling; reasoning is always on (default `max`).
13
+ * - **`qwen38-27b`** (Provisioned Throughput) — 262,144-token context window; image input (inline base64); tool calling; reasoning on by default (`none` turns it off).
11
14
  */
12
15
  export declare const V2ModelIds: {
13
16
  readonly Sarvam105B: "sarvam-105b";
17
+ readonly Sarvam105BConversations: "sarvam-105b-conversations";
14
18
  readonly Glm53: "glm5.3";
15
19
  readonly Glm53Flash: "glm5.3-flash";
16
20
  readonly Gemma4: "gemma4";
17
21
  readonly Deepseekv41Flash1017: "deepseekv4.1-flash-1017";
22
+ readonly Deepseekv41Flash1017Interactive: "deepseekv4.1-flash-1017-interactive";
18
23
  readonly Deepseekv4Pro: "deepseekv4-pro";
19
24
  readonly KimiK31Pro: "kimi-k3.1-pro";
25
+ readonly Qwen3827B: "qwen38-27b";
20
26
  };
21
27
  export type V2ModelIds = (typeof V2ModelIds)[keyof typeof V2ModelIds];
@@ -6,19 +6,25 @@ exports.V2ModelIds = void 0;
6
6
  * Model IDs accepted by `POST /v2/chat/completions` (case-sensitive; unknown ids return `404`):
7
7
  *
8
8
  * - **`sarvam-105b`** — Sarvam flagship chat model.
9
+ * - **`sarvam-105b-conversations`** — Sarvam 105B tuned for real-time dialogue and voice agents; 32,000-token context window.
9
10
  * - **`glm5.3`** — 1,048,576-token context window; visible reasoning in `reasoning_content` (billed as completion tokens); tool calling; max request size 10,485,760 bytes; `n` must be `1`.
10
11
  * - **`glm5.3-flash`** — 1,048,576-token context window, image input (inline base64), tool calling, and reasoning (on by default).
11
12
  * - **`gemma4`** — 262,144-token context window, image input (inline base64), tool calling, and optional reasoning (`reasoning_effort`).
12
- * - **`deepseekv4.1-flash-1017`** — 1,048,576-token context window, tool calling, and reasoning.
13
+ * - **`deepseekv4.1-flash-1017`** — 1,048,576-token context window, image input (inline base64), tool calling, and reasoning.
14
+ * - **`deepseekv4.1-flash-1017-interactive`** — the same model on a high-throughput deployment: more than 200 output tokens per second, against about 80; same capabilities as `deepseekv4.1-flash-1017`.
13
15
  * - **`deepseekv4-pro`** (Provisioned Throughput) — 1,000,000-token context window; text input only; tool calling; reasoning (default `high`; `none` turns it off).
14
- * - **`kimi-k3.1-pro`** (Provisioned Throughput) — 1,048,576-token context window; image input and short video clips (inline base64); tool calling; reasoning is always on (default `max`).
16
+ * - **`kimi-k3.1-pro`** (Provisioned Throughput) — 1,048,576-token context window; image and video input (inline base64); tool calling; reasoning is always on (default `max`).
17
+ * - **`qwen38-27b`** (Provisioned Throughput) — 262,144-token context window; image input (inline base64); tool calling; reasoning on by default (`none` turns it off).
15
18
  */
16
19
  exports.V2ModelIds = {
17
20
  Sarvam105B: "sarvam-105b",
21
+ Sarvam105BConversations: "sarvam-105b-conversations",
18
22
  Glm53: "glm5.3",
19
23
  Glm53Flash: "glm5.3-flash",
20
24
  Gemma4: "gemma4",
21
25
  Deepseekv41Flash1017: "deepseekv4.1-flash-1017",
26
+ Deepseekv41Flash1017Interactive: "deepseekv4.1-flash-1017-interactive",
22
27
  Deepseekv4Pro: "deepseekv4-pro",
23
28
  KimiK31Pro: "kimi-k3.1-pro",
29
+ Qwen3827B: "qwen38-27b",
24
30
  };
@@ -1 +1 @@
1
- export declare const SDK_VERSION = "1.1.16-alpha.3";
1
+ export declare const SDK_VERSION = "1.1.17";
@@ -1,4 +1,4 @@
1
1
  "use strict";
2
2
  Object.defineProperty(exports, "__esModule", { value: true });
3
3
  exports.SDK_VERSION = void 0;
4
- exports.SDK_VERSION = "1.1.16-alpha.3";
4
+ exports.SDK_VERSION = "1.1.17";
@@ -6,8 +6,8 @@ export function normalizeClientOptions(options) {
6
6
  const headers = mergeHeaders({
7
7
  "X-Fern-Language": "JavaScript",
8
8
  "X-Fern-SDK-Name": "sarvamai",
9
- "X-Fern-SDK-Version": "1.1.16-alpha.3",
10
- "User-Agent": "sarvamai/1.1.16-alpha.3",
9
+ "X-Fern-SDK-Version": "1.1.17",
10
+ "User-Agent": "sarvamai/1.1.17",
11
11
  "X-Fern-Runtime": core.RUNTIME.type,
12
12
  "X-Fern-Runtime-Version": core.RUNTIME.version,
13
13
  }, options === null || options === void 0 ? void 0 : options.headers);
@@ -305,8 +305,12 @@ export class ChatClient {
305
305
  throw new SarvamAI.BadRequestError(body, rawResponse);
306
306
  case 401:
307
307
  throw new SarvamAI.UnauthorizedError(body, rawResponse);
308
+ case 402:
309
+ throw new SarvamAI.PaymentRequiredError(body, rawResponse);
308
310
  case 403:
309
311
  throw new SarvamAI.ForbiddenError(body, rawResponse);
312
+ case 404:
313
+ throw new SarvamAI.NotFoundError(body, rawResponse);
310
314
  case 413:
311
315
  throw new SarvamAI.ContentTooLargeError(body, rawResponse);
312
316
  case 422:
@@ -315,8 +319,12 @@ export class ChatClient {
315
319
  throw new SarvamAI.TooManyRequestsError(body, rawResponse);
316
320
  case 500:
317
321
  throw new SarvamAI.InternalServerError(body, rawResponse);
322
+ case 502:
323
+ throw new SarvamAI.BadGatewayError(body, rawResponse);
318
324
  case 503:
319
325
  throw new SarvamAI.ServiceUnavailableError(body, rawResponse);
326
+ case 504:
327
+ throw new SarvamAI.GatewayTimeoutError(body, rawResponse);
320
328
  default:
321
329
  throw new errors.SarvamAIError({
322
330
  statusCode,
@@ -20,7 +20,7 @@ export interface CreateChatCompletionV2Request {
20
20
  temperature?: number;
21
21
  /** Nucleus sampling: greater than `0` and at most `1`. On `glm5.3`, `top_p: 0` is rejected. Prefer changing either `top_p` or `temperature`, not both. */
22
22
  top_p?: number;
23
- /** On reasoning models (`glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`): **`low`**, **`high`**, or **`max`**. Default when omitted is **`max`**, except **`high`** on `glm5.3-flash`, `deepseekv4.1-flash-1017`, and `deepseekv4-pro`, and no reasoning on `gemma4`. **`none`** turns thinking off on `deepseekv4.1-flash-1017` and `deepseekv4-pro`. `kimi-k3.1-pro` always thinks. Trace in `message.reasoning_content`; billed as completion tokens and counts toward `max_tokens`. On `gemma4`, reasoning is off unless this is set (`low`, `high`, or `max`). */
23
+ /** On reasoning models: **`low`**, **`high`**, or **`max`** on every reasoning model, and **`medium`** on `sarvam-105b`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, `kimi-k3.1-pro`, and `qwen38-27b` (`kimi-k3.1-pro` runs `medium` as `high`). Default when omitted is **`max`** on `glm5.3` and `kimi-k3.1-pro`, **`high`** on `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro`, and no reasoning on `gemma4`; `qwen38-27b` thinks by default. **`none`** turns thinking off on `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`. `kimi-k3.1-pro` always thinks: `none`, `minimal`, and `xhigh` return `400`. Trace in `message.reasoning_content`; billed as completion tokens and counts toward `max_tokens`. On `gemma4`, reasoning is off unless this is set (`low`, `high`, or `max`). */
24
24
  reasoning_effort?: SarvamAI.ReasoningEffort;
25
25
  /** Maximum tokens to generate in the completion, including reasoning tokens on reasoning models. Default `2048`. Set explicitly when you need longer answers — a low limit can yield `finish_reason: "length"` with little or no visible `content` if reasoning consumes the budget. */
26
26
  max_tokens?: number;
@@ -59,8 +59,8 @@ export interface CreateChatCompletionV2Request {
59
59
  parallel_tool_calls?: boolean;
60
60
  /** Only valid with `stream: true` (otherwise `400`). Must be a top-level field, not inside `extra_body`. Use `{"include_usage": true}` to receive token usage in the final SSE chunk. */
61
61
  stream_options?: SarvamAI.ChatStreamOptions;
62
- /** If `true`, return log probabilities for output tokens (use with `top_logprobs`). Supported on **`sarvam-105b`**, **`gemma4`**, **`glm5.3-flash`**, **`deepseekv4.1-flash-1017`**, **`deepseekv4-pro`**, and **`kimi-k3.1-pro`**. Not supported on **`glm5.3`** (`400` if enabled). */
62
+ /** If `true`, return log probabilities for output tokens (use with `top_logprobs`). Supported on **`sarvam-105b`**, **`gemma4`**, **`glm5.3-flash`**, **`deepseekv4.1-flash-1017`**, **`deepseekv4.1-flash-1017-interactive`**, **`deepseekv4-pro`**, and **`qwen38-27b`**. Not supported on **`glm5.3`** or **`kimi-k3.1-pro`** (`400` if enabled). */
63
63
  logprobs?: boolean;
64
- /** How many top token log probabilities to return at each position. Takes effect only when `logprobs` is `true`; without it, some models ignore the field and others return **`400`**. Same model support as `logprobs`: works on `sarvam-105b`, `gemma4`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro`; on `glm5.3`, omit entirely — do not send `0`, since many SDKs add it by default. */
64
+ /** How many top token log probabilities to return at each position. Takes effect only when `logprobs` is `true`; without it, some models ignore the field and others return **`400`**. Same model support as `logprobs`: works on `sarvam-105b`, `gemma4`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`; on `glm5.3` and `kimi-k3.1-pro`, omit entirely — do not send `0`, since many SDKs add it by default. */
65
65
  top_logprobs?: number;
66
66
  }
@@ -11,17 +11,25 @@ export declare class DecisionsClient {
11
11
  protected readonly _options: NormalizedClientOptionsWithAuth<DecisionsClient.Options>;
12
12
  constructor(options?: DecisionsClient.Options);
13
13
  /**
14
- * Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse. The order of `choice` options does not affect the answer.
14
+ * Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse.
15
+ *
16
+ * **Models:** `sarvam-prism` (serverless) and `clef` (Provisioned Throughput; also reads images).
15
17
  *
16
18
  * Requires `api-subscription-key`, or `Authorization: Bearer <key>`.
17
19
  *
18
20
  * **Question types:** `noul` (yes/no), `choice` (one of your options) and `score` (a level on your ordered scale).
19
21
  *
20
- * **Repeatability:** probabilities can vary slightly between identical requests; the top answer is stable.
22
+ * **Repeatability:** on `sarvam-prism`, probabilities can vary slightly between identical requests and the top answer is stable; the order of `choice` options does not affect the answer. On `clef`, identical requests return identical probabilities.
23
+ *
24
+ * **Images:** `clef` reads base64 images in `images`; image URLs return `400`. `sarvam-prism` refuses `images` (image URLs return `400`, inline images `422`).
25
+ *
26
+ * **Not supported:** streaming. `temperature` returns `422` on both models. `sarvam-prism` also refuses `max_tokens` and `stream` with `422`; `clef` accepts and ignores them. Unknown keys inside a question return `422`.
21
27
  *
22
- * **Not supported:** streaming, and images on `sarvam-prism` (`clef` accepts base64 images). Generation fields such as `temperature`, `max_tokens` and `stream` are refused with `422`. On `sarvam-prism`, `images` is refused too (image URLs return `400`, inline images `422`); `clef` reads base64 images and refuses image URLs with `400`. Unknown keys inside a question also return `422`.
28
+ * **Limits:**
29
+ * - `sarvam-prism`: 32,768 tokens per question (state + question + options), 2–255 options per `choice`, and 2–10 levels per `score`. If any one question breaks a limit, the whole request fails with `422`.
30
+ * - `clef`: 65,536 tokens per request (state, every question and its options, and images), 1–255 options per `choice`, and 1–10 levels per `score`. A longer request fails with `400`.
23
31
  *
24
- * **Limits:** 32,768 tokens per question (state + question + options), 2–255 options per `choice` and 2–10 levels per `score`. There is no cap on the number of questions. The state is not truncated, and if any one question breaks a limit the whole request fails with `422`.
32
+ * There is no cap on the number of questions, and the state is never truncated.
25
33
  *
26
34
  * **Timeout:** each request must finish within 60 seconds. A longer request fails with `503 model_overloaded`; split very large batches across several requests before retrying.
27
35
  *
@@ -20,17 +20,25 @@ export class DecisionsClient {
20
20
  this._options = normalizeClientOptionsWithAuth(options);
21
21
  }
22
22
  /**
23
- * Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse. The order of `choice` options does not affect the answer.
23
+ * Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse.
24
+ *
25
+ * **Models:** `sarvam-prism` (serverless) and `clef` (Provisioned Throughput; also reads images).
24
26
  *
25
27
  * Requires `api-subscription-key`, or `Authorization: Bearer <key>`.
26
28
  *
27
29
  * **Question types:** `noul` (yes/no), `choice` (one of your options) and `score` (a level on your ordered scale).
28
30
  *
29
- * **Repeatability:** probabilities can vary slightly between identical requests; the top answer is stable.
31
+ * **Repeatability:** on `sarvam-prism`, probabilities can vary slightly between identical requests and the top answer is stable; the order of `choice` options does not affect the answer. On `clef`, identical requests return identical probabilities.
32
+ *
33
+ * **Images:** `clef` reads base64 images in `images`; image URLs return `400`. `sarvam-prism` refuses `images` (image URLs return `400`, inline images `422`).
34
+ *
35
+ * **Not supported:** streaming. `temperature` returns `422` on both models. `sarvam-prism` also refuses `max_tokens` and `stream` with `422`; `clef` accepts and ignores them. Unknown keys inside a question return `422`.
30
36
  *
31
- * **Not supported:** streaming, and images on `sarvam-prism` (`clef` accepts base64 images). Generation fields such as `temperature`, `max_tokens` and `stream` are refused with `422`. On `sarvam-prism`, `images` is refused too (image URLs return `400`, inline images `422`); `clef` reads base64 images and refuses image URLs with `400`. Unknown keys inside a question also return `422`.
37
+ * **Limits:**
38
+ * - `sarvam-prism`: 32,768 tokens per question (state + question + options), 2–255 options per `choice`, and 2–10 levels per `score`. If any one question breaks a limit, the whole request fails with `422`.
39
+ * - `clef`: 65,536 tokens per request (state, every question and its options, and images), 1–255 options per `choice`, and 1–10 levels per `score`. A longer request fails with `400`.
32
40
  *
33
- * **Limits:** 32,768 tokens per question (state + question + options), 2–255 options per `choice` and 2–10 levels per `score`. There is no cap on the number of questions. The state is not truncated, and if any one question breaks a limit the whole request fails with `422`.
41
+ * There is no cap on the number of questions, and the state is never truncated.
34
42
  *
35
43
  * **Timeout:** each request must finish within 60 seconds. A longer request fails with `503 model_overloaded`; split very large batches across several requests before retrying.
36
44
  *
@@ -139,6 +139,8 @@ export class ResponsesClient {
139
139
  throw new SarvamAI.BadRequestError(body, rawResponse);
140
140
  case 401:
141
141
  throw new SarvamAI.UnauthorizedError(body, rawResponse);
142
+ case 402:
143
+ throw new SarvamAI.PaymentRequiredError(body, rawResponse);
142
144
  case 403:
143
145
  throw new SarvamAI.ForbiddenError(body, rawResponse);
144
146
  case 404:
@@ -185,6 +187,8 @@ export class ResponsesClient {
185
187
  throw new SarvamAI.BadRequestError(_response.error.body, _response.rawResponse);
186
188
  case 401:
187
189
  throw new SarvamAI.UnauthorizedError(_response.error.body, _response.rawResponse);
190
+ case 402:
191
+ throw new SarvamAI.PaymentRequiredError(_response.error.body, _response.rawResponse);
188
192
  case 403:
189
193
  throw new SarvamAI.ForbiddenError(_response.error.body, _response.rawResponse);
190
194
  case 404:
@@ -19,7 +19,7 @@ export interface ResponsesCreateRequest {
19
19
  tools?: SarvamAI.ResponseFunctionTool[];
20
20
  /** Forwarded exactly as written: `none`, `auto`, `required`, `{ "type": "function", "name": "..." }`, or another object. */
21
21
  tool_choice?: ResponsesCreateRequest.ToolChoice;
22
- /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro`, `low`, `high`, or `max` all work directly. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
22
+ /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, the DeepSeek models, `kimi-k3.1-pro`, and `qwen38-27b`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro`, `low`, `high`, `max`, or `none` (thinking off); on `kimi-k3.1-pro`, `low`, `medium`, `high`, or `max` (thinking is always on); on `qwen38-27b`, `none` turns thinking off. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
23
23
  reasoning?: SarvamAI.ResponseReasoning;
24
24
  /** Merged into the outbound body. Must not overwrite any modelled field (`400`). `max_tool_calls` and `prompt` are `400` here and in `extra_body` — not supported by any serving deployment. */
25
25
  extra_body?: Record<string, unknown>;
@@ -48,6 +48,10 @@ export interface ResponsesCreateRequest {
48
48
  background?: boolean;
49
49
  /** `400 invalid_request_error` — not supported by any serving deployment, top-level or via `extra_body`. */
50
50
  max_tool_calls?: number;
51
+ /** How many of the most likely tokens to return at each output position, with their log probabilities. Send it with `include: ["message.output_text.logprobs"]`; without `include`, `output[].content[].logprobs` comes back empty. `0` returns no logprobs. Supported on `sarvam-105b`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`; `glm5.3`, `gemma4`, and `kimi-k3.1-pro` return `400`. */
52
+ top_logprobs?: number;
53
+ /** Extra data to add to the response. `message.output_text.logprobs` returns token log probabilities (use with `top_logprobs`); `reasoning.encrypted_content` is accepted. An unknown value returns `400`. */
54
+ include?: string[];
51
55
  /** Accepts any additional properties */
52
56
  [key: string]: any;
53
57
  }
@@ -19,7 +19,7 @@ export interface ResponsesCreateStreamRequest {
19
19
  tools?: SarvamAI.ResponseFunctionTool[];
20
20
  /** Forwarded exactly as written: `none`, `auto`, `required`, `{ "type": "function", "name": "..." }`, or another object. */
21
21
  tool_choice?: ResponsesCreateStreamRequest.ToolChoice;
22
- /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro`, `low`, `high`, or `max` all work directly. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
22
+ /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, the DeepSeek models, `kimi-k3.1-pro`, and `qwen38-27b`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro`, `low`, `high`, `max`, or `none` (thinking off); on `kimi-k3.1-pro`, `low`, `medium`, `high`, or `max` (thinking is always on); on `qwen38-27b`, `none` turns thinking off. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
23
23
  reasoning?: SarvamAI.ResponseReasoning;
24
24
  /** Merged into the outbound body. Must not overwrite any modelled field (`400`). `max_tool_calls` and `prompt` are `400` here and in `extra_body` — not supported by any serving deployment. */
25
25
  extra_body?: Record<string, unknown>;
@@ -48,6 +48,10 @@ export interface ResponsesCreateStreamRequest {
48
48
  background?: boolean;
49
49
  /** `400 invalid_request_error` — not supported by any serving deployment, top-level or via `extra_body`. */
50
50
  max_tool_calls?: number;
51
+ /** How many of the most likely tokens to return at each output position, with their log probabilities. Send it with `include: ["message.output_text.logprobs"]`; without `include`, `output[].content[].logprobs` comes back empty. `0` returns no logprobs. Supported on `sarvam-105b`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`; `glm5.3`, `gemma4`, and `kimi-k3.1-pro` return `400`. */
52
+ top_logprobs?: number;
53
+ /** Extra data to add to the response. `message.output_text.logprobs` returns token log probabilities (use with `top_logprobs`); `reasoning.encrypted_content` is accepted. An unknown value returns `400`. */
54
+ include?: string[];
51
55
  /** Accepts any additional properties */
52
56
  [key: string]: any;
53
57
  }
@@ -9,7 +9,7 @@ export interface Choice {
9
9
  finish_reason: SarvamAI.FinishReason;
10
10
  /** The index of the choice in the list of choices. */
11
11
  index: number;
12
- /** Token log probabilities when supported (`sarvam-105b`, `gemma4`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`). Always `null` on `glm5.3`, which doesn't support this field. */
12
+ /** Token log probabilities when supported (`sarvam-105b`, `gemma4`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, `qwen38-27b`). Not supported on `glm5.3` or `kimi-k3.1-pro`, which return `400` when `logprobs` is set. */
13
13
  logprobs?: Record<string, unknown> | undefined;
14
14
  message: SarvamAI.ChatCompletionResponseMessage;
15
15
  }
@@ -14,7 +14,7 @@ export interface CreateResponseV2Request {
14
14
  tools?: SarvamAI.ResponseFunctionTool[] | undefined;
15
15
  /** Forwarded exactly as written: `none`, `auto`, `required`, `{ "type": "function", "name": "..." }`, or another object. */
16
16
  tool_choice?: CreateResponseV2Request.ToolChoice | undefined;
17
- /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro`, `low`, `high`, or `max` all work directly. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
17
+ /** Optional reasoning config for `glm5.3`, `glm5.3-flash`, `gemma4`, the DeepSeek models, `kimi-k3.1-pro`, and `qwen38-27b`. **`effort`:** on `glm5.3` and `glm5.3-flash`, `low` or `high` only (the literal string `max` returns `400`; omitting it gives `max` on `glm5.3` and `high` on `glm5.3-flash`); on `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro`, `low`, `high`, `max`, or `none` (thinking off); on `kimi-k3.1-pro`, `low`, `medium`, `high`, or `max` (thinking is always on); on `qwen38-27b`, `none` turns thinking off. The trace is in the reasoning item's `content[0].text`; `summary` is filled only when `summary` is set. Optional **`summary`:** `auto` or `detailed`. */
18
18
  reasoning?: SarvamAI.ResponseReasoning | undefined;
19
19
  /** Merged into the outbound body. Must not overwrite any modelled field (`400`). `max_tool_calls` and `prompt` are `400` here and in `extra_body` — not supported by any serving deployment. */
20
20
  extra_body?: Record<string, unknown> | undefined;
@@ -43,6 +43,10 @@ export interface CreateResponseV2Request {
43
43
  background?: boolean | undefined;
44
44
  /** `400 invalid_request_error` — not supported by any serving deployment, top-level or via `extra_body`. */
45
45
  max_tool_calls?: number | undefined;
46
+ /** How many of the most likely tokens to return at each output position, with their log probabilities. Send it with `include: ["message.output_text.logprobs"]`; without `include`, `output[].content[].logprobs` comes back empty. `0` returns no logprobs. Supported on `sarvam-105b`, `glm5.3-flash`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, and `qwen38-27b`; `glm5.3`, `gemma4`, and `kimi-k3.1-pro` return `400`. */
47
+ top_logprobs?: number | undefined;
48
+ /** Extra data to add to the response. `message.output_text.logprobs` returns token log probabilities (use with `top_logprobs`); `reasoning.encrypted_content` is accepted. An unknown value returns `400`. */
49
+ include?: string[] | undefined;
46
50
  /** Accepts any additional properties */
47
51
  [key: string]: any;
48
52
  }
@@ -3,4 +3,14 @@ export interface ModelObject {
3
3
  object?: "model" | undefined;
4
4
  created: number;
5
5
  owned_by: string;
6
+ /** How the model is offered. `"provisioned"`: Provisioned Throughput, enabled per organisation (without access, `deepseekv4-pro` returns `400` and the others return `404` and aren't listed). `"standard"`: serverless, any API key can call it. */
7
+ access?: ModelObject.Access | undefined;
8
+ }
9
+ export declare namespace ModelObject {
10
+ /** How the model is offered. `"provisioned"`: Provisioned Throughput, enabled per organisation (without access, `deepseekv4-pro` returns `400` and the others return `404` and aren't listed). `"standard"`: serverless, any API key can call it. */
11
+ const Access: {
12
+ readonly Standard: "standard";
13
+ readonly Provisioned: "provisioned";
14
+ };
15
+ type Access = (typeof Access)[keyof typeof Access];
6
16
  }
@@ -1,2 +1,9 @@
1
1
  // This file was auto-generated by Fern from our API Definition.
2
- export {};
2
+ export var ModelObject;
3
+ (function (ModelObject) {
4
+ /** How the model is offered. `"provisioned"`: Provisioned Throughput, enabled per organisation (without access, `deepseekv4-pro` returns `400` and the others return `404` and aren't listed). `"standard"`: serverless, any API key can call it. */
5
+ ModelObject.Access = {
6
+ Standard: "standard",
7
+ Provisioned: "provisioned",
8
+ };
9
+ })(ModelObject || (ModelObject = {}));
@@ -1,7 +1,8 @@
1
- /** Reasoning depth on `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`: `low`, `high`, or `max`. Default when omitted is `max` (`high` on `glm5.3-flash`, `deepseekv4.1-flash-1017`, and `deepseekv4-pro`; no reasoning on `gemma4`). `none` turns thinking off on `deepseekv4.1-flash-1017` and `deepseekv4-pro`; other models reject it or keep thinking. */
1
+ /** Reasoning depth. `low`, `high`, and `max` work on every reasoning model; `medium` also works on `sarvam-105b`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, `kimi-k3.1-pro`, and `qwen38-27b`. Default when omitted is `max` on `glm5.3` and `kimi-k3.1-pro`, `high` on `glm5.3-flash` and the DeepSeek models, and no reasoning on `gemma4`. `none` turns thinking off on the DeepSeek models and `qwen38-27b`; `kimi-k3.1-pro` refuses it, and other models reject it or keep thinking. */
2
2
  export declare const ReasoningEffort: {
3
3
  readonly None: "none";
4
4
  readonly Low: "low";
5
+ readonly Medium: "medium";
5
6
  readonly High: "high";
6
7
  readonly Max: "max";
7
8
  };
@@ -1,8 +1,9 @@
1
1
  // This file was auto-generated by Fern from our API Definition.
2
- /** Reasoning depth on `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3.1-pro`: `low`, `high`, or `max`. Default when omitted is `max` (`high` on `glm5.3-flash`, `deepseekv4.1-flash-1017`, and `deepseekv4-pro`; no reasoning on `gemma4`). `none` turns thinking off on `deepseekv4.1-flash-1017` and `deepseekv4-pro`; other models reject it or keep thinking. */
2
+ /** Reasoning depth. `low`, `high`, and `max` work on every reasoning model; `medium` also works on `sarvam-105b`, `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, `deepseekv4-pro`, `kimi-k3.1-pro`, and `qwen38-27b`. Default when omitted is `max` on `glm5.3` and `kimi-k3.1-pro`, `high` on `glm5.3-flash` and the DeepSeek models, and no reasoning on `gemma4`. `none` turns thinking off on the DeepSeek models and `qwen38-27b`; `kimi-k3.1-pro` refuses it, and other models reject it or keep thinking. */
3
3
  export const ReasoningEffort = {
4
4
  None: "none",
5
5
  Low: "low",
6
+ Medium: "medium",
6
7
  High: "high",
7
8
  Max: "max",
8
9
  };
@@ -1,6 +1,6 @@
1
1
  import type * as SarvamAI from "../index.mjs";
2
2
  export interface ResponseReasoning {
3
- /** `low`, `high`, or `max`. On **`glm5.3`** and **`glm5.3-flash`**, sending the literal string `max` returns `400`; omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. **`deepseekv4.1-flash-1017`**, **`deepseekv4-pro`**, and **`kimi-k3.1-pro`** accept `max` directly on this route. `none` turns thinking off on `deepseekv4-pro`. `gemma4` accepts `low` or `high`; omitting it means no reasoning. */
3
+ /** `low`, `high`, or `max`, plus `medium` and `none` on some models. On **`glm5.3`** and **`glm5.3-flash`**, sending the literal string `max` returns `400`; omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. **`deepseekv4.1-flash-1017`**, **`deepseekv4.1-flash-1017-interactive`**, and **`deepseekv4-pro`** accept `max` directly, and `none` turns thinking off. **`kimi-k3.1-pro`** accepts `low`, `medium`, `high`, or `max` and always thinks. **`qwen38-27b`** thinks by default; `none` turns it off. `gemma4` accepts `low` or `high`; omitting it means no reasoning. */
4
4
  effort?: SarvamAI.ResponseReasoningEffort | undefined;
5
5
  /** `auto` or `detailed`. Without it, the reasoning item's `summary` is empty; read the trace from `content[0].text`. */
6
6
  summary?: ResponseReasoning.Summary | undefined;
@@ -1,7 +1,8 @@
1
- /** Reasoning effort on the Responses route. `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro` accept all three values directly. `glm5.3` and `glm5.3-flash` accept `low` and `high`; the literal string `max` returns `400` on them. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. `none` turns thinking off on `deepseekv4-pro`. `gemma4` accepts `low` or `high`; omitting the field means no reasoning. */
1
+ /** Reasoning effort on the Responses route. `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro` accept `low`, `high`, and `max` directly, and `none` turns thinking off. `kimi-k3.1-pro` accepts `low`, `medium`, `high`, and `max` and always thinks. `qwen38-27b` thinks by default; `none` turns it off. `glm5.3` and `glm5.3-flash` accept `low` and `high`; the literal string `max` returns `400` on them. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. `gemma4` accepts `low` or `high`; omitting the field means no reasoning. */
2
2
  export declare const ResponseReasoningEffort: {
3
3
  readonly None: "none";
4
4
  readonly Low: "low";
5
+ readonly Medium: "medium";
5
6
  readonly High: "high";
6
7
  readonly Max: "max";
7
8
  };
@@ -1,8 +1,9 @@
1
1
  // This file was auto-generated by Fern from our API Definition.
2
- /** Reasoning effort on the Responses route. `deepseekv4.1-flash-1017`, `deepseekv4-pro`, and `kimi-k3.1-pro` accept all three values directly. `glm5.3` and `glm5.3-flash` accept `low` and `high`; the literal string `max` returns `400` on them. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. `none` turns thinking off on `deepseekv4-pro`. `gemma4` accepts `low` or `high`; omitting the field means no reasoning. */
2
+ /** Reasoning effort on the Responses route. `deepseekv4.1-flash-1017`, `deepseekv4.1-flash-1017-interactive`, and `deepseekv4-pro` accept `low`, `high`, and `max` directly, and `none` turns thinking off. `kimi-k3.1-pro` accepts `low`, `medium`, `high`, and `max` and always thinks. `qwen38-27b` thinks by default; `none` turns it off. `glm5.3` and `glm5.3-flash` accept `low` and `high`; the literal string `max` returns `400` on them. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. `gemma4` accepts `low` or `high`; omitting the field means no reasoning. */
3
3
  export const ResponseReasoningEffort = {
4
4
  None: "none",
5
5
  Low: "low",
6
+ Medium: "medium",
6
7
  High: "high",
7
8
  Max: "max",
8
9
  };
@@ -2,20 +2,26 @@
2
2
  * Model IDs accepted by `POST /v2/responses` (case-sensitive; unknown ids return `404`):
3
3
  *
4
4
  * - **`sarvam-105b`** — Sarvam flagship model.
5
+ * - **`sarvam-105b-conversations`** — Sarvam 105B tuned for real-time dialogue.
5
6
  * - **`glm5.3`** — 1,048,576-token context; reasoning in `output` as a `reasoning` item; flat function tools; max request size 10,485,760 bytes.
6
7
  * - **`glm5.3-flash`** — long context, vision (inline base64 images), tools, and reasoning; `reasoning.effort` accepts `low` or `high` only, as on `glm5.3`.
7
8
  * - **`gemma4`** — 262,144-token context, vision (inline base64 images), tools, and optional reasoning (`reasoning.effort` `low` or `high`).
8
- * - **`deepseekv4.1-flash-1017`** — long context, tools, and reasoning.
9
+ * - **`deepseekv4.1-flash-1017`** — long context, vision (inline base64 images), tools, and reasoning.
10
+ * - **`deepseekv4.1-flash-1017-interactive`** — same capabilities as `deepseekv4.1-flash-1017`, on a high-throughput deployment (more than 200 output tokens per second).
9
11
  * - **`deepseekv4-pro`** (Provisioned Throughput) — long context, text input, tools, and reasoning (default `high`).
10
12
  * - **`kimi-k3.1-pro`** (Provisioned Throughput) — long context, image input (inline base64; no video on this route), tools, and reasoning (always on).
13
+ * - **`qwen38-27b`** (Provisioned Throughput) — 262,144-token context, image input (inline base64), tools, and reasoning (on by default).
11
14
  */
12
15
  export declare const ResponsesModelIds: {
13
16
  readonly Sarvam105B: "sarvam-105b";
17
+ readonly Sarvam105BConversations: "sarvam-105b-conversations";
14
18
  readonly Glm53: "glm5.3";
15
19
  readonly Glm53Flash: "glm5.3-flash";
16
20
  readonly Gemma4: "gemma4";
17
21
  readonly Deepseekv41Flash1017: "deepseekv4.1-flash-1017";
22
+ readonly Deepseekv41Flash1017Interactive: "deepseekv4.1-flash-1017-interactive";
18
23
  readonly Deepseekv4Pro: "deepseekv4-pro";
19
24
  readonly KimiK31Pro: "kimi-k3.1-pro";
25
+ readonly Qwen3827B: "qwen38-27b";
20
26
  };
21
27
  export type ResponsesModelIds = (typeof ResponsesModelIds)[keyof typeof ResponsesModelIds];
@@ -3,19 +3,25 @@
3
3
  * Model IDs accepted by `POST /v2/responses` (case-sensitive; unknown ids return `404`):
4
4
  *
5
5
  * - **`sarvam-105b`** — Sarvam flagship model.
6
+ * - **`sarvam-105b-conversations`** — Sarvam 105B tuned for real-time dialogue.
6
7
  * - **`glm5.3`** — 1,048,576-token context; reasoning in `output` as a `reasoning` item; flat function tools; max request size 10,485,760 bytes.
7
8
  * - **`glm5.3-flash`** — long context, vision (inline base64 images), tools, and reasoning; `reasoning.effort` accepts `low` or `high` only, as on `glm5.3`.
8
9
  * - **`gemma4`** — 262,144-token context, vision (inline base64 images), tools, and optional reasoning (`reasoning.effort` `low` or `high`).
9
- * - **`deepseekv4.1-flash-1017`** — long context, tools, and reasoning.
10
+ * - **`deepseekv4.1-flash-1017`** — long context, vision (inline base64 images), tools, and reasoning.
11
+ * - **`deepseekv4.1-flash-1017-interactive`** — same capabilities as `deepseekv4.1-flash-1017`, on a high-throughput deployment (more than 200 output tokens per second).
10
12
  * - **`deepseekv4-pro`** (Provisioned Throughput) — long context, text input, tools, and reasoning (default `high`).
11
13
  * - **`kimi-k3.1-pro`** (Provisioned Throughput) — long context, image input (inline base64; no video on this route), tools, and reasoning (always on).
14
+ * - **`qwen38-27b`** (Provisioned Throughput) — 262,144-token context, image input (inline base64), tools, and reasoning (on by default).
12
15
  */
13
16
  export const ResponsesModelIds = {
14
17
  Sarvam105B: "sarvam-105b",
18
+ Sarvam105BConversations: "sarvam-105b-conversations",
15
19
  Glm53: "glm5.3",
16
20
  Glm53Flash: "glm5.3-flash",
17
21
  Gemma4: "gemma4",
18
22
  Deepseekv41Flash1017: "deepseekv4.1-flash-1017",
23
+ Deepseekv41Flash1017Interactive: "deepseekv4.1-flash-1017-interactive",
19
24
  Deepseekv4Pro: "deepseekv4-pro",
20
25
  KimiK31Pro: "kimi-k3.1-pro",
26
+ Qwen3827B: "qwen38-27b",
21
27
  };
@@ -2,20 +2,26 @@
2
2
  * Model IDs accepted by `POST /v2/chat/completions` (case-sensitive; unknown ids return `404`):
3
3
  *
4
4
  * - **`sarvam-105b`** — Sarvam flagship chat model.
5
+ * - **`sarvam-105b-conversations`** — Sarvam 105B tuned for real-time dialogue and voice agents; 32,000-token context window.
5
6
  * - **`glm5.3`** — 1,048,576-token context window; visible reasoning in `reasoning_content` (billed as completion tokens); tool calling; max request size 10,485,760 bytes; `n` must be `1`.
6
7
  * - **`glm5.3-flash`** — 1,048,576-token context window, image input (inline base64), tool calling, and reasoning (on by default).
7
8
  * - **`gemma4`** — 262,144-token context window, image input (inline base64), tool calling, and optional reasoning (`reasoning_effort`).
8
- * - **`deepseekv4.1-flash-1017`** — 1,048,576-token context window, tool calling, and reasoning.
9
+ * - **`deepseekv4.1-flash-1017`** — 1,048,576-token context window, image input (inline base64), tool calling, and reasoning.
10
+ * - **`deepseekv4.1-flash-1017-interactive`** — the same model on a high-throughput deployment: more than 200 output tokens per second, against about 80; same capabilities as `deepseekv4.1-flash-1017`.
9
11
  * - **`deepseekv4-pro`** (Provisioned Throughput) — 1,000,000-token context window; text input only; tool calling; reasoning (default `high`; `none` turns it off).
10
- * - **`kimi-k3.1-pro`** (Provisioned Throughput) — 1,048,576-token context window; image input and short video clips (inline base64); tool calling; reasoning is always on (default `max`).
12
+ * - **`kimi-k3.1-pro`** (Provisioned Throughput) — 1,048,576-token context window; image and video input (inline base64); tool calling; reasoning is always on (default `max`).
13
+ * - **`qwen38-27b`** (Provisioned Throughput) — 262,144-token context window; image input (inline base64); tool calling; reasoning on by default (`none` turns it off).
11
14
  */
12
15
  export declare const V2ModelIds: {
13
16
  readonly Sarvam105B: "sarvam-105b";
17
+ readonly Sarvam105BConversations: "sarvam-105b-conversations";
14
18
  readonly Glm53: "glm5.3";
15
19
  readonly Glm53Flash: "glm5.3-flash";
16
20
  readonly Gemma4: "gemma4";
17
21
  readonly Deepseekv41Flash1017: "deepseekv4.1-flash-1017";
22
+ readonly Deepseekv41Flash1017Interactive: "deepseekv4.1-flash-1017-interactive";
18
23
  readonly Deepseekv4Pro: "deepseekv4-pro";
19
24
  readonly KimiK31Pro: "kimi-k3.1-pro";
25
+ readonly Qwen3827B: "qwen38-27b";
20
26
  };
21
27
  export type V2ModelIds = (typeof V2ModelIds)[keyof typeof V2ModelIds];
@@ -3,19 +3,25 @@
3
3
  * Model IDs accepted by `POST /v2/chat/completions` (case-sensitive; unknown ids return `404`):
4
4
  *
5
5
  * - **`sarvam-105b`** — Sarvam flagship chat model.
6
+ * - **`sarvam-105b-conversations`** — Sarvam 105B tuned for real-time dialogue and voice agents; 32,000-token context window.
6
7
  * - **`glm5.3`** — 1,048,576-token context window; visible reasoning in `reasoning_content` (billed as completion tokens); tool calling; max request size 10,485,760 bytes; `n` must be `1`.
7
8
  * - **`glm5.3-flash`** — 1,048,576-token context window, image input (inline base64), tool calling, and reasoning (on by default).
8
9
  * - **`gemma4`** — 262,144-token context window, image input (inline base64), tool calling, and optional reasoning (`reasoning_effort`).
9
- * - **`deepseekv4.1-flash-1017`** — 1,048,576-token context window, tool calling, and reasoning.
10
+ * - **`deepseekv4.1-flash-1017`** — 1,048,576-token context window, image input (inline base64), tool calling, and reasoning.
11
+ * - **`deepseekv4.1-flash-1017-interactive`** — the same model on a high-throughput deployment: more than 200 output tokens per second, against about 80; same capabilities as `deepseekv4.1-flash-1017`.
10
12
  * - **`deepseekv4-pro`** (Provisioned Throughput) — 1,000,000-token context window; text input only; tool calling; reasoning (default `high`; `none` turns it off).
11
- * - **`kimi-k3.1-pro`** (Provisioned Throughput) — 1,048,576-token context window; image input and short video clips (inline base64); tool calling; reasoning is always on (default `max`).
13
+ * - **`kimi-k3.1-pro`** (Provisioned Throughput) — 1,048,576-token context window; image and video input (inline base64); tool calling; reasoning is always on (default `max`).
14
+ * - **`qwen38-27b`** (Provisioned Throughput) — 262,144-token context window; image input (inline base64); tool calling; reasoning on by default (`none` turns it off).
12
15
  */
13
16
  export const V2ModelIds = {
14
17
  Sarvam105B: "sarvam-105b",
18
+ Sarvam105BConversations: "sarvam-105b-conversations",
15
19
  Glm53: "glm5.3",
16
20
  Glm53Flash: "glm5.3-flash",
17
21
  Gemma4: "gemma4",
18
22
  Deepseekv41Flash1017: "deepseekv4.1-flash-1017",
23
+ Deepseekv41Flash1017Interactive: "deepseekv4.1-flash-1017-interactive",
19
24
  Deepseekv4Pro: "deepseekv4-pro",
20
25
  KimiK31Pro: "kimi-k3.1-pro",
26
+ Qwen3827B: "qwen38-27b",
21
27
  };
@@ -1 +1 @@
1
- export declare const SDK_VERSION = "1.1.16-alpha.3";
1
+ export declare const SDK_VERSION = "1.1.17";
@@ -1 +1 @@
1
- export const SDK_VERSION = "1.1.16-alpha.3";
1
+ export const SDK_VERSION = "1.1.17";
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "sarvamai",
3
- "version": "1.1.16-alpha.3",
3
+ "version": "1.1.17",
4
4
  "private": false,
5
5
  "repository": {
6
6
  "type": "git",
package/reference.md CHANGED
@@ -3874,7 +3874,7 @@ await client.docAi.getCapabilities();
3874
3874
 
3875
3875
  Generate a model reply in the OpenAI Responses format.
3876
3876
 
3877
- **Models:** `sarvam-105b`, `glm5.3`, `gemma4`, `deepseekv4.1-flash-1017`, `glm5.3-flash`, `deepseekv4-pro`, `kimi-k3.1-pro` (ids are case-sensitive).
3877
+ **Models:** `sarvam-105b`, `sarvam-105b-conversations`, `glm5.3`, `gemma4`, `deepseekv4.1-flash-1017`, `glm5.3-flash`, `deepseekv4.1-flash-1017-interactive`, and, on Provisioned Throughput, `deepseekv4-pro`, `kimi-k3.1-pro`, `qwen38-27b` (ids are case-sensitive).
3878
3878
 
3879
3879
  **Each call is independent.** Send the full conversation in `input` every time — nothing is stored for you. Leave out `store` or set it to `false` (`true` returns **`400`**). You cannot load a past response by id (`GET /v2/responses/{id}` is always **`404`**).
3880
3880
 
@@ -3953,7 +3953,7 @@ for await (const item of response) {
3953
3953
 
3954
3954
  Generate a model reply in the OpenAI Responses format.
3955
3955
 
3956
- **Models:** `sarvam-105b`, `glm5.3`, `gemma4`, `deepseekv4.1-flash-1017`, `glm5.3-flash`, `deepseekv4-pro`, `kimi-k3.1-pro` (ids are case-sensitive).
3956
+ **Models:** `sarvam-105b`, `sarvam-105b-conversations`, `glm5.3`, `gemma4`, `deepseekv4.1-flash-1017`, `glm5.3-flash`, `deepseekv4.1-flash-1017-interactive`, and, on Provisioned Throughput, `deepseekv4-pro`, `kimi-k3.1-pro`, `qwen38-27b` (ids are case-sensitive).
3957
3957
 
3958
3958
  **Each call is independent.** Send the full conversation in `input` every time — nothing is stored for you. Leave out `store` or set it to `false` (`true` returns **`400`**). You cannot load a past response by id (`GET /v2/responses/{id}` is always **`404`**).
3959
3959
 
@@ -4147,17 +4147,25 @@ await client.chatCompletion.getModelV2("model_id");
4147
4147
  <dl>
4148
4148
  <dd>
4149
4149
 
4150
- Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse. The order of `choice` options does not affect the answer.
4150
+ Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse.
4151
+
4152
+ **Models:** `sarvam-prism` (serverless) and `clef` (Provisioned Throughput; also reads images).
4151
4153
 
4152
4154
  Requires `api-subscription-key`, or `Authorization: Bearer <key>`.
4153
4155
 
4154
4156
  **Question types:** `noul` (yes/no), `choice` (one of your options) and `score` (a level on your ordered scale).
4155
4157
 
4156
- **Repeatability:** probabilities can vary slightly between identical requests; the top answer is stable.
4158
+ **Repeatability:** on `sarvam-prism`, probabilities can vary slightly between identical requests and the top answer is stable; the order of `choice` options does not affect the answer. On `clef`, identical requests return identical probabilities.
4159
+
4160
+ **Images:** `clef` reads base64 images in `images`; image URLs return `400`. `sarvam-prism` refuses `images` (image URLs return `400`, inline images `422`).
4161
+
4162
+ **Not supported:** streaming. `temperature` returns `422` on both models. `sarvam-prism` also refuses `max_tokens` and `stream` with `422`; `clef` accepts and ignores them. Unknown keys inside a question return `422`.
4157
4163
 
4158
- **Not supported:** streaming, and images on `sarvam-prism` (`clef` accepts base64 images). Generation fields such as `temperature`, `max_tokens` and `stream` are refused with `422`. On `sarvam-prism`, `images` is refused too (image URLs return `400`, inline images `422`); `clef` reads base64 images and refuses image URLs with `400`. Unknown keys inside a question also return `422`.
4164
+ **Limits:**
4165
+ - `sarvam-prism`: 32,768 tokens per question (state + question + options), 2–255 options per `choice`, and 2–10 levels per `score`. If any one question breaks a limit, the whole request fails with `422`.
4166
+ - `clef`: 65,536 tokens per request (state, every question and its options, and images), 1–255 options per `choice`, and 1–10 levels per `score`. A longer request fails with `400`.
4159
4167
 
4160
- **Limits:** 32,768 tokens per question (state + question + options), 2–255 options per `choice` and 2–10 levels per `score`. There is no cap on the number of questions. The state is not truncated, and if any one question breaks a limit the whole request fails with `422`.
4168
+ There is no cap on the number of questions, and the state is never truncated.
4161
4169
 
4162
4170
  **Timeout:** each request must finish within 60 seconds. A longer request fails with `503 model_overloaded`; split very large batches across several requests before retrying.
4163
4171