@juspay/neurolink 12.39.0 → 12.40.0

package/dist/providers/catalog/vultr-inference.json

@@ -0,0 +1,220 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "vultr-inference",
+    "displayName": "Vultr Inference",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.vultrinference.com/v1"
+    },
+    "models": {
+        "default": "glm-5.2",
+        "fallbacks": [
+            "deepseek-v4.1-flash",
+            "qwen3.8-flash-next",
+            "mimo-v2.6-flash-rl"
+        ],
+        "defaultContextWindow": 131072,
+        "defaultMaxOutputTokens": 32768,
+        "catalog": {
+            "glm-5.2": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.75,
+                    "output": 3
+                },
+                "vision": true,
+                "status": "production",
+                "description": "NeuroLink default. Roster entry named \"GLM 5.2\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"GLM-5.2 is Z.ai's flagship model for long-horizon tasks, sustaining agentic work across a solid 1M-token context.\" Roster fields: text input max_context_length 1,048,576 tokens; output max_length 1,048,576 tokens; prompt price $0.75 and completion price $3 per MTok (cost_usd 0.00000075 and 0.000003 per token); streaming true; tools listed among the supported parameters; input modalities text, image. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "deepseek-v4.1-flash": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.15,
+                    "output": 0.6
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Fallback. Roster entry named \"DeepSeek V4.1 Flash\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"DeepSeek V4.1 Flash is a mixture-of-experts model with 552B backbone parameters, a one-million-token context window, and support for reasoning and tool calling.\" Roster fields: text input max_context_length 1,048,576 tokens; output max_length 1,048,576 tokens; prompt price $0.15 and completion price $0.6 per MTok (cost_usd 0.00000015 and 0.0000006 per token); streaming true; tools listed among the supported parameters; input modalities text, image. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "qwen3.8-flash-next": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 262144,
+                "pricingPerMTok": {
+                    "input": 0.1,
+                    "output": 0.2
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Fallback. Roster entry named \"Qwen 3.8 Flash Next\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"Qwen3.8 Flash Next with NVFP4 weights, a 262,144-token context window, reasoning, and tool calling.\" Roster fields: text input max_context_length 262,144 tokens; output max_length 262,144 tokens; prompt price $0.1 and completion price $0.2 per MTok (cost_usd 0.0000001 and 0.0000002 per token); streaming true; tools listed among the supported parameters; input modalities text, image, video. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "mimo-v2.6-flash-rl": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.1,
+                    "output": 0.25
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Fallback. Roster entry named \"MiMo V2.6 Flash RL\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"MiMo V2.6 Flash RL is Xiaomi's efficiency-balanced mixture-of-experts model with 309B total and 15B active parameters, hybrid sliding-window attention, and a 1M-token context.\" \"Supports reasoning and tool calling, served at FP8.\" Roster fields: text input max_context_length 1,048,576 tokens; output max_length 1,048,576 tokens; prompt price $0.1 and completion price $0.25 per MTok (cost_usd 0.0000001 and 0.00000025 per token); streaming true; tools listed among the supported parameters; input modalities text, image, audio, video. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "deepseek-v4-flash-0731": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.1,
+                    "output": 0.25
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Roster entry named \"DeepSeek V4 Flash 0731\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"DeepSeek's official V4 Flash release with substantially enhanced agentic capabilities and a built-in speculative decoding module.\" Roster fields: text input max_context_length 1,048,576 tokens; output max_length 1,048,576 tokens; prompt price $0.1 and completion price $0.25 per MTok (cost_usd 0.0000001 and 0.00000025 per token); streaming true; tools listed among the supported parameters; input modalities text, image. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "glm-5.3": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.75,
+                    "output": 3
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Roster entry named \"GLM 5.3\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"Served at NVFP4 across a 1M-token context.\" Roster fields: text input max_context_length 1,048,576 tokens; output max_length 1,048,576 tokens; prompt price $0.75 and completion price $3 per MTok (cost_usd 0.00000075 and 0.000003 per token); streaming true; tools listed among the supported parameters; input modalities text, image. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "glm-5.3-flash": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.1,
+                    "output": 0.35
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Roster entry named \"GLM 5.3 Flash\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"Z.ai's first natively multimodal GLM-5 model: a 320B-parameter MoE with ~18B active per token, combining sparse and linear attention for cheap long-context serving across a 1M-token window.\" Roster fields: text input max_context_length 1,048,576 tokens; output max_length 1,048,576 tokens; prompt price $0.1 and completion price $0.35 per MTok (cost_usd 0.0000001 and 0.00000035 per token); streaming true; tools listed among the supported parameters; input modalities text, image, video. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "glm-5.x-menthol": {
+                "contextWindow": 202752,
+                "maxOutputTokens": 202752,
+                "pricingPerMTok": {
+                    "input": 0.4,
+                    "output": 1.75
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Roster entry named \"GLM 5\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"Z.ai's GLM-5, a 744B-parameter MoE flagship for agentic reasoning and coding, served in FP8.\" Roster fields: text input max_context_length 202,752 tokens; output max_length 202,752 tokens; prompt price $0.4 and completion price $1.75 per MTok (cost_usd 0.0000004 and 0.00000175 per token); streaming true; tools listed among the supported parameters; input modalities text, image. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "laguna-s-2.1": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.09,
+                    "output": 0.18
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Roster entry named \"Laguna S 2.1\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"Laguna S 2.1 is a 118B total parameter Mixture-of-Experts model with 8B activated parameters per token, designed for agentic coding and long-horizon work.\" Roster fields: text input max_context_length 1,048,576 tokens; output max_length 1,048,576 tokens; prompt price $0.09 and completion price $0.18 per MTok (cost_usd 0.00000009 and 0.00000018 per token); streaming true; tools listed among the supported parameters; input modalities text, image. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "mimo-v2.6-pro-rl": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.4,
+                    "output": 0.8
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Roster entry named \"MiMo V2.6 Pro RL\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"MiMo V2.6 Pro RL is Xiaomi's mixture-of-experts model with 1.02T total and 42B active parameters, hybrid sliding-window attention, and a 1M-token context.\" \"Supports reasoning and tool calling, served at FP8.\" Roster fields: text input max_context_length 1,048,576 tokens; output max_length 1,048,576 tokens; prompt price $0.4 and completion price $0.8 per MTok (cost_usd 0.0000004 and 0.0000008 per token); streaming true; tools listed among the supported parameters; input modalities text, image, audio, video. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            },
+            "minimax-m3": {
+                "contextWindow": 524288,
+                "maxOutputTokens": 524288,
+                "pricingPerMTok": {
+                    "input": 0.2,
+                    "output": 0.9
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Roster entry named \"MiniMax M3\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"MiniMax M3 is a natively multimodal MoE model (~428B parameters, ~23B active) with a 1M-token context.\" Roster fields: text input max_context_length 524,288 tokens; output max_length 524,288 tokens; prompt price $0.2 and completion price $0.9 per MTok (cost_usd 0.0000002 and 0.0000009 per token); streaming true; tools listed among the supported parameters; input modalities text, image, video. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent. The roster description text says a 1M-token context while the max_context_length field says 524,288; contextWindow uses the field."
+            },
+            "muse-glimmer-30b": {
+                "contextWindow": 131072,
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 0.25,
+                    "output": 1
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Roster entry named \"Muse Glimmer 30B\" on the unauthenticated GET https://api.vultrinference.com/v1/models (2026-09-29, 24 entries). Roster description: \"Muse Glimmer 30B is a dense multimodal model for reasoning, coding, and tool use, with a 131072-token context.\" Roster fields: text input max_context_length 131,072 tokens; output max_length 131,072 tokens; prompt price $0.25 and completion price $1 per MTok (cost_usd 0.00000025 and 0.000001 per token); streaming true; tools listed among the supported parameters; input modalities text, image, video. Vision is true because the roster lists an image input modality (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif); no image request was sent."
+            }
+        },
+        "topModels": [
+            "glm-5.2",
+            "deepseek-v4.1-flash",
+            "qwen3.8-flash-next",
+            "mimo-v2.6-flash-rl",
+            "deepseek-v4-flash-0731",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "glm-5.x-menthol",
+            "laguna-s-2.1",
+            "mimo-v2.6-pro-rl",
+            "minimax-m3",
+            "muse-glimmer-30b"
+        ]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": "model-dependent",
+        "toolsWithStreaming": true,
+        "structuredOutput": false,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": false
+    },
+    "errorRules": [
+        {
+            "status": 401,
+            "class": "authentication",
+            "message": "Vultr Inference rejected the request with HTTP 401 (Unauthorized). Check {apiKeyEnvVar}; the steps to retrieve a key are at {setupUrl}"
+        },
+        {
+            "status": 429,
+            "class": "rate-limit",
+            "message": "Vultr Inference reports HTTP 429 (Rate limit exceeded)."
+        }
+    ],
+    "setup": {
+        "url": "https://docs.vultr.com/products/compute/serverless-inference/provisioning",
+        "apiKeyFormat": null,
+        "billingPolicy": "no-free-tier",
+        "instructions": [
+            "1. Visit: https://docs.vultr.com/products/compute/serverless-inference/provisioning and follow the Vultr Console steps: Products, Serverless, Inference, Add Serverless Inference, then provide a Label, \"acknowledge the list of supported models and the charges note\", and click Add Serverless Inference",
+            "2. Open the new subscription's management page and copy the API key from its overview page (https://docs.vultr.com/products/compute/serverless-inference/management/connection: \"API key is necessary to send any request to the inference endpoints\")",
+            "3. Charges: the provisioning page above lists the charges note among the items to acknowledge, and the FAQ (https://docs.vultr.com/products/compute/serverless-inference/faq) says the subscription's Usage tab shows \"details on your current token usage, overage, and any associated costs\"",
+            "4. Set {apiKeyEnvVar} in your .env file (the authentication example on https://api.vultrinference.com/ writes ${VULTR_API_KEY}, which NeuroLink does not read)"
+        ],
+        "description": "Chat completions endpoint at https://api.vultrinference.com/v1 (https://api.vultrinference.com/ documents POST /chat/completions with bearer-token authentication and says \"Responses follow OpenRouter's chat completion shape whichever engine serves the model\"; the FAQ at https://docs.vultr.com/products/compute/serverless-inference/faq says to \"replace your current inference API URL (such as OpenAI's base API URL) with Vultr’s API URL\"). The chat-completions request message schema at https://api.vultrinference.com/ lists role (system, assistant, user, tool, developer) and content (type string, \"The contents of the message.\"); the tool call id description says \"Send it back unchanged as the `tool_call_id` of the tool result.\" The roster route is public (\"Most API routes require a bearer token. The public `/v1/models` route does not.\") and lists 24 entries, 15 with a text output; the catalog holds 12 chat models. billingPolicy no-free-tier is the schema's closest value (the schema has no unknown value) and is not a Vultr statement. Each model's status production is a value the schema requires and is not a Vultr statement. models.defaultContextWindow (131,072) and models.defaultMaxOutputTokens (32,768) are placeholders for ids outside the catalog that the vendor does not publish: 131,072 is the context the roster lists for muse-glimmer-30b and 32,768 is the default the chat-completions page shows for max_completion_tokens. errorRules match the documented statuses 401 and 429; the 401 body was observed on unauthenticated GETs of /v1/models/all, /v1/usage and /v1/models/glm-5.2 ({\"message\":\"No API key provided\"}); no POST to /chat/completions was sent."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-29",
+            "status": 200,
+            "method": "unauthenticated GET https://api.vultrinference.com/v1/models — 24 models; credential-free, no key used"
+        },
+        "authProbe": {
+            "date": "2026-09-29",
+            "status": 401,
+            "method": "unauthenticated GET https://api.vultrinference.com/v1/models/all with no credentials — 401 {\"message\":\"No API key provided\"} (the same body on GET /v1/usage and GET /v1/models/glm-5.2); the body carries no error code; GET https://api.vultrinference.com/v1/chat/completions answered 404 (the documented method is POST and none was sent); credential-free, no key used"
+        },
+        "billingProbe": {
+            "date": "2026-09-29",
+            "method": "public pages https://docs.vultr.com/products/compute/serverless-inference/provisioning (Console steps: \"acknowledge the list of supported models and the charges note\") and https://docs.vultr.com/products/compute/serverless-inference/faq (Usage tab: \"details on your current token usage, overage, and any associated costs\"); no free-tier statement read; credential-free"
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1860"
+    }
+}

package/dist/providers/catalog/wafer.json

@@ -0,0 +1,165 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "wafer",
+    "displayName": "Wafer",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://pass.wafer.ai/v1"
+    },
+    "models": {
+        "default": "GLM-5.3",
+        "fallbacks": ["DeepSeek-V4-Pro", "Kimi-K3", "GLM-5.2"],
+        "defaultContextWindow": 128000,
+        "defaultMaxOutputTokens": 16384,
+        "catalog": {
+            "GLM-5.3": {
+                "contextWindow": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.29,
+                    "output": 4.4,
+                    "cachedInput": 0.2359
+                },
+                "vision": false,
+                "status": "production",
+                "description": "NeuroLink default. The model named in the curl request example on https://docs.wafer.ai/serverless and on https://docs.wafer.ai/serverless/zero-data-retention. Unauthenticated GET https://pass.wafer.ai/v1/models roster (2026-09-29, 8 models), wafer.description: \"GLM-5.3 — Z.ai's flagship GLM-5.3 MoE (391B, NVFP4), self-hosted on the Wafer fleet with a 1M-token context window and frontier coding/agentic performance. Available serverless.\" Roster values: context_length 1,048,576; input_cents_per_million 29, output_cents_per_million 440 ($0.29 / $4.40 per MTok input/output); cache_read_microcents_per_million 23,590,000 ($0.2359 per MTok cache-read); capabilities.vision false; chat_completions supported, tools, streaming, json_object and json_schema true."
+            },
+            "DeepSeek-V4-Pro": {
+                "contextWindow": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.4,
+                    "output": 4.2,
+                    "cachedInput": 0.32
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Fallback. Unauthenticated GET https://pass.wafer.ai/v1/models roster (2026-09-29, 8 models), wafer.description: \"DeepSeek-V4-Pro -- DeepSeek's V4 Pro MoE, self-hosted on the Wafer fleet with a 1M-token context window, reasoning-effort control and tool calling. Available serverless.\" Roster values: context_length 1,048,576; input_cents_per_million 40, output_cents_per_million 420 ($0.40 / $4.20 per MTok input/output); cache_read_microcents_per_million 32,000,000 ($0.32 per MTok cache-read); capabilities.vision false; chat_completions supported, tools, streaming, json_object and json_schema true."
+            },
+            "Kimi-K3": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 1.44,
+                    "output": 9,
+                    "cachedInput": 0.3
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Fallback. Unauthenticated GET https://pass.wafer.ai/v1/models roster (2026-09-29, 8 models), wafer.description: \"Kimi K3 sparse MoE model, self-hosted on the Wafer fleet. Available serverless with a 1M-token context window and strong coding/agentic performance.\" Roster values: context_length 1,048,576; max_output_tokens 131,072; input_cents_per_million 144, output_cents_per_million 900 ($1.44 / $9.00 per MTok input/output); cache_read_microcents_per_million 30,000,000 ($0.3 per MTok cache-read); capabilities.vision true and supports_vision true; not probed with an image request; chat_completions supported, tools, streaming, json_object and json_schema true."
+            },
+            "GLM-5.2": {
+                "contextWindow": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.32,
+                    "output": 4.4,
+                    "cachedInput": 0.2599
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Fallback. Unauthenticated GET https://pass.wafer.ai/v1/models roster (2026-09-29, 8 models), wafer.description: \"GLM-5.2 — Z.ai's GLM-5.2 MoE (NVFP4), self-hosted on the Wafer fleet with a 1M-token context window, strong coding and reasoning, and tool calling. Available serverless.\" Roster values: context_length 1,048,576; input_cents_per_million 32, output_cents_per_million 440 ($0.32 / $4.40 per MTok input/output); cache_read_microcents_per_million 25,990,000 ($0.2599 per MTok cache-read); capabilities.vision false; chat_completions supported, tools, streaming, json_object and json_schema true."
+            },
+            "Qwen3.8-27B": {
+                "contextWindow": 262144,
+                "pricingPerMTok": {
+                    "input": 0.04,
+                    "output": 4.4,
+                    "cachedInput": 0.0342
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Unauthenticated GET https://pass.wafer.ai/v1/models roster (2026-09-29, 8 models), wafer.description: \"Qwen3.8-27B -- Alibaba's dense 27B vision-language model, self-hosted on the Wafer fleet with a 262K-token context window, image and video input, reasoning-effort control and tool calling. Available serverless.\" Roster values: context_length 262,144; input_cents_per_million 4, output_cents_per_million 440 ($0.04 / $4.40 per MTok input/output); cache_read_microcents_per_million 3,420,000 ($0.0342 per MTok cache-read); capabilities.vision true and supports_vision true; not probed with an image request; chat_completions supported, tools, streaming, json_object and json_schema true."
+            },
+            "GLM-5.3-Flash": {
+                "contextWindow": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.75,
+                    "output": 0.5,
+                    "cachedInput": 0.03
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Unauthenticated GET https://pass.wafer.ai/v1/models roster (2026-09-29, 8 models), wafer.description: \"GLM-5.3-Flash — Z.ai's fast GLM-5.3 MoE variant, self-hosted on the Wafer fleet with a 1M-token context window and strong coding/agentic performance. Available serverless to gateway partners.\" Roster values: context_length 1,048,576; input_cents_per_million 75, output_cents_per_million 50 ($0.75 / $0.50 per MTok input/output); cache_read_microcents_per_million 3,000,000 ($0.03 per MTok cache-read); capabilities.vision true and supports_vision true; not probed with an image request; chat_completions supported, tools, streaming, json_object and json_schema true."
+            },
+            "DeepSeek-V4.1-Flash": {
+                "contextWindow": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.02,
+                    "output": 0.7,
+                    "cachedInput": 0.016
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Roster-listed. Unauthenticated GET https://pass.wafer.ai/v1/models roster (2026-09-29, 8 models), wafer.description: \"DeepSeek-V4.1-Flash — DeepSeek's V4.1 Flash MoE (552B total, 8B active for prefill / 16B for decode, 1M-token context), self-hosted on the Wafer fleet with reasoning-effort control and tool calling. Available serverless.\" Roster values: context_length 1,048,576; input_cents_per_million 2, output_cents_per_million 70 ($0.02 / $0.70 per MTok input/output); cache_read_microcents_per_million 1,600,000 ($0.016 per MTok cache-read); capabilities.vision true and supports_vision true; not probed with an image request; chat_completions supported, tools, streaming, json_object and json_schema true."
+            },
+            "DeepSeek-V4-Flash-0731-Fast": {
+                "contextWindow": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.07,
+                    "output": 0.35,
+                    "cachedInput": 0.0531
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Roster-listed. Unauthenticated GET https://pass.wafer.ai/v1/models roster (2026-09-29, 8 models), wafer.description: \"The same model served for high TPS.\" Roster values: context_length 1,048,576; input_cents_per_million 7, output_cents_per_million 35 ($0.07 / $0.35 per MTok input/output); cache_read_microcents_per_million 5,310,000 ($0.0531 per MTok cache-read); capabilities.vision false; chat_completions supported, tools, streaming, json_object and json_schema true."
+            }
+        },
+        "topModels": [
+            "GLM-5.3",
+            "DeepSeek-V4-Pro",
+            "Kimi-K3",
+            "GLM-5.2",
+            "Qwen3.8-27B",
+            "GLM-5.3-Flash",
+            "DeepSeek-V4.1-Flash",
+            "DeepSeek-V4-Flash-0731-Fast"
+        ],
+        "visionModel": "Qwen3.8-27B"
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": true,
+        "toolsWithStreaming": true,
+        "structuredOutput": true,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": false
+    },
+    "errorRules": [
+        {
+            "status": 401,
+            "class": "authentication",
+            "message": "Wafer returned HTTP 401 authentication_error. Check {apiKeyEnvVar}. Get one at {setupUrl}"
+        }
+    ],
+    "setup": {
+        "url": "https://app.wafer.ai",
+        "apiKeyFormat": null,
+        "billingPolicy": "no-free-tier",
+        "instructions": [
+            "1. Visit: https://app.wafer.ai and sign in — the docs say \"Sign in at app.wafer.ai to add credits and create an API key. Store your key securely.\" (https://docs.wafer.ai/serverless)",
+            "2. Add credits and create an API key — the docs say \"Add credits, create an API key, and pay per token.\" (https://docs.wafer.ai/serverless)",
+            "3. Set {apiKeyEnvVar} in your .env file (the docs' Authentication row shows the Authorization: Bearer header)"
+        ],
+        "description": "OpenAI-compatible endpoint at https://pass.wafer.ai/v1 (https://docs.wafer.ai/serverless: \"Access Wafer models through an OpenAI-compatible API.\"), with the 8 models the unauthenticated GET /models roster listed on 2026-09-29 in models.catalog. An unauthenticated GET of https://pass.wafer.ai/v1/chat/completions returned HTTP 401 with error.type authentication_error and error.code missing_api_key (2026-09-29), so errorRules match status 401. models.defaultContextWindow (128000) and models.defaultMaxOutputTokens (16384) are placeholders the vendor does not publish. billingPolicy is recorded as no-free-tier because the schema has no unknown value; the vendor's stated flow is \"Add credits, create an API key, and pay per token.\""
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-29",
+            "status": 200,
+            "method": "unauthenticated GET https://pass.wafer.ai/v1/models — 8 models; credential-free, no key used"
+        },
+        "authProbe": {
+            "date": "2026-09-29",
+            "status": 401,
+            "code": "missing_api_key",
+            "method": "unauthenticated GET https://pass.wafer.ai/v1/chat/completions — error.type authentication_error; credential-free, no key used"
+        },
+        "billingProbe": {
+            "date": "2026-09-29",
+            "method": "public page https://docs.wafer.ai/serverless — \"Add credits, create an API key, and pay per token.\"; credential-free"
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1860"
+    }
+}