@aliou/pi-neuralwatt 0.16.1 → 0.16.3

package/extensions/provider/api/openai-completions.ts

@@ -1,3 +1,4 @@
+import type { StreamOptions } from "@earendil-works/pi-ai";
 import { stream, streamSimple } from "@earendil-works/pi-ai/compat";
 import {
   NEURALWATT_BASE_URL,
@@ -8,9 +9,56 @@ import type { NeuralwattModel } from "../models/catalog";
 import type { AnyStreamSimple } from "../stream-simple";
 import type { NeuralwattApiHandler } from "./types";
 
+type OpenAiCompletionsBody = {
+  messages?: Array<Record<string, unknown>>;
+  [key: string]: unknown;
+};
+
+/**
+ * Neuralwatt streams chain-of-thought in the `reasoning` field, and pi-ai
+ * replays prior thinking under the field name it recorded from the stream —
+ * also `reasoning`. The served chat templates only render `reasoning_content`
+ * (verified against the Kimi K3 template), so replayed thinking never reaches
+ * the model. Rename the replayed field on the wire. This is a move, not a
+ * copy: sending an empty `reasoning_content` next to a populated `reasoning`
+ * makes the gateway prefer the empty field and silently drops the replay.
+ */
+function makeReasoningReplayInjector(
+  upstream?: StreamOptions["onPayload"],
+): NonNullable<StreamOptions["onPayload"]> {
+  return async (payload, model) => {
+    const next = await upstream?.(payload, model);
+    const body = (next !== undefined ? next : payload) as OpenAiCompletionsBody;
+    const messages = body.messages;
+    if (!Array.isArray(messages)) return body;
+
+    return {
+      ...body,
+      messages: messages.map((message) => {
+        if (message?.role !== "assistant" || !("reasoning" in message)) {
+          return message;
+        }
+        const { reasoning, ...rest } = message;
+        // A pre-set non-empty reasoning_content wins; drop the duplicate.
+        return typeof rest.reasoning_content === "string" &&
+          rest.reasoning_content.length > 0
+          ? rest
+          : { ...rest, reasoning_content: reasoning };
+      }),
+    };
+  };
+}
+
 export function createOpenAiCompletionsApi(options?: {
   streamSimple?: AnyStreamSimple;
 }): NeuralwattApiHandler {
+  const withReasoningReplay = (options?: {
+    onPayload?: StreamOptions["onPayload"];
+  }) => ({
+    ...options,
+    onPayload: makeReasoningReplayInjector(options?.onPayload),
+  });
+
   return {
     stampModels: (models: NeuralwattModel[]) =>
       models.map((model) => {
@@ -24,7 +72,12 @@ export function createOpenAiCompletionsApi(options?: {
         };
       }),
     stream: (model, context, streamOptions) =>
-      stream(model, context, streamOptions as never),
-    streamSimple: options?.streamSimple ?? streamSimple,
+      stream(model, context, withReasoningReplay(streamOptions) as never),
+    streamSimple: (model, context, simpleOptions) =>
+      (options?.streamSimple ?? streamSimple)(
+        model,
+        context,
+        withReasoningReplay(simpleOptions) as never,
+      ),
   };
 }

package/extensions/provider/models/public-models.ts

@@ -116,6 +116,62 @@ const KIMI_K2_7_CODE: NeuralwattModelFamily = {
   },
 };
 
+// Xiaomi. MiMo-V2.6-Pro only has two reasoning behaviors: `high` (default)
+// and `none`; the API aliases every other effort onto those, with `minimal`
+// landing on `none`.
+const MIMO_V2_6_PRO: NeuralwattModelFamily = {
+  cost: { input: 0.87, output: 1.74, cacheRead: 0.036 },
+  vision: true,
+  reasoningMetadata: {
+    supported_efforts: ["high", "none"],
+    mandatory: false,
+    effort_aliases: {
+      max: "high",
+      xhigh: "high",
+      medium: "high",
+      low: "high",
+      minimal: "none",
+    },
+  },
+};
+
+// Neuralwatt tracking alias: the current large model, today Kimi K3. The
+// alias moves to newer models over time, so it snapshots K3's contract
+// instead of sharing the KIMI_K3 family.
+const NW_LARGE: NeuralwattModelFamily = {
+  cost: { input: 3, output: 15, cacheRead: 0.3 },
+  vision: true,
+  reasoningMetadata: {
+    supported_efforts: ["max", "high", "low", "none"],
+    mandatory: false,
+    effort_aliases: { xhigh: "max", medium: "high", minimal: "low" },
+  },
+};
+
+// Neuralwatt tracking alias: the current cheap, fast model, today DeepSeek
+// V4.1 Flash. Same moving-alias caveat as NW_LARGE.
+const NW_FLASH: NeuralwattModelFamily = {
+  cost: { input: 0.15, output: 0.6, cacheRead: 0.015 },
+  vision: true,
+  reasoningMetadata: {
+    supported_efforts: ["max", "xhigh", "high", "low", "none"],
+    mandatory: false,
+    effort_aliases: { medium: "high", minimal: "low" },
+  },
+};
+
+// Neuralwatt tracking alias: the current small model, today Qwen3.8-27B.
+// Same moving-alias caveat as NW_LARGE.
+const NW_SMALL: NeuralwattModelFamily = {
+  cost: { input: 0.45, output: 3.2, cacheRead: 0.25 },
+  vision: true,
+  reasoningMetadata: {
+    supported_efforts: ["xhigh", "medium", "low", "none"],
+    mandatory: false,
+    effort_aliases: { max: "xhigh", high: "xhigh", minimal: "low" },
+  },
+};
+
 // Qwen. Qwen3.6 35B only advertises `high` and `none`.
 const QWEN_3_6_35B: NeuralwattModelFamily = {
   cost: { input: 0.29, output: 1.15, cacheRead: 0.029 },
@@ -192,6 +248,13 @@ const FAMILIES: [NeuralwattModelFamily, NeuralwattVariantSpec[]][] = [
         reasoning: true,
         costMultiplier: 0.65,
       },
+      {
+        id: "deepseek-v4.1-flash-speed",
+        name: "DeepSeek V4.1 Flash (Speed)",
+        contextWindow: 1048560,
+        maxOutputTokens: 393216,
+        reasoning: true,
+      },
     ],
   ],
   [
@@ -302,27 +365,99 @@ const FAMILIES: [NeuralwattModelFamily, NeuralwattVariantSpec[]][] = [
       },
     ],
   ],
+  [
+    MIMO_V2_6_PRO,
+    [
+      {
+        id: "mimo-v2.6-pro",
+        name: "MiMo-V2.6-Pro",
+        contextWindow: 1048560,
+        maxOutputTokens: 393216,
+        reasoning: true,
+      },
+    ],
+  ],
+  [
+    NW_LARGE,
+    [
+      {
+        id: "nw-large",
+        name: "Neuralwatt Large",
+        contextWindow: 1048560,
+        maxOutputTokens: null,
+        reasoning: true,
+      },
+      {
+        id: "nw-large-flex",
+        name: "Neuralwatt Large (flex)",
+        contextWindow: 1048560,
+        maxOutputTokens: null,
+        reasoning: true,
+        costMultiplier: 0.65,
+      },
+    ],
+  ],
+  [
+    NW_FLASH,
+    [
+      {
+        id: "nw-flash",
+        name: "Neuralwatt Flash",
+        contextWindow: 1048560,
+        maxOutputTokens: 393216,
+        reasoning: true,
+      },
+      {
+        id: "nw-flash-flex",
+        name: "Neuralwatt Flash (flex)",
+        contextWindow: 1048560,
+        maxOutputTokens: 393216,
+        reasoning: true,
+        costMultiplier: 0.65,
+      },
+    ],
+  ],
+  [
+    NW_SMALL,
+    [
+      {
+        id: "nw-small",
+        name: "Neuralwatt Small",
+        contextWindow: 262128,
+        maxOutputTokens: 131072,
+        reasoning: true,
+      },
+      {
+        id: "nw-small-flex",
+        name: "Neuralwatt Small (flex)",
+        contextWindow: 262128,
+        maxOutputTokens: 131072,
+        reasoning: true,
+        costMultiplier: 0.65,
+      },
+    ],
+  ],
   [
     QWEN_3_6_35B,
     [
       {
         id: "qwen3.6-35b",
         name: "Qwen3.6 35B",
-        contextWindow: 131056,
+        contextWindow: 262128,
         maxOutputTokens: null,
         reasoning: true,
       },
       {
         id: "qwen3.6-35b-fast",
         name: "Qwen3.6 35B Fast",
-        contextWindow: 131056,
+        contextWindow: 262128,
         maxOutputTokens: null,
         reasoning: false,
       },
       {
         id: "qwen3.6-35b-flex",
         name: "Qwen3.6 35B (flex)",
-        contextWindow: 131056,
+        contextWindow: 262128,
         maxOutputTokens: null,
         reasoning: true,
         costMultiplier: 0.65,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aliou/pi-neuralwatt",
-  "version": "0.16.1",
+  "version": "0.16.3",
   "license": "MIT",
   "type": "module",
   "private": false,