@aliou/pi-neuralwatt 0.10.4 → 0.10.5

package/extensions/provider/models/hidden.ts

@@ -3,35 +3,32 @@ import { fetchNeuralwattModels } from "../../../src/lib/neuralwatt-api";
 import type { NeuralwattApiModel } from "../../../src/types/models-api";
 import { NEURALWATT_MODELS } from "./public-models";
 
-// Hidden aliases that work for authorized accounts but are omitted from the
-// authenticated /v1/models response. Keep these gated by includeHiddenModels.
+// Hidden models available to authorized accounts but omitted from the public
+// /v1/models response. Keep these gated by includeHiddenModels and hardcode
+// early-access entries so they remain available from the offline catalog.
 // Move an entry to public-models.ts once Neuralwatt advertises it publicly.
 export const HIDDEN_NEURALWATT_MODELS: ProviderModelConfig[] = [
-  // DeepSeek V4 Flash Canary - early-access 284B/13B-active MoE served on B200s.
-  // Neuralwatt omits it from the authenticated catalog. Context and runtime
-  // capabilities were verified directly; pricing follows DeepSeek's upstream
-  // rates until Neuralwatt publishes model metadata.
+  // Kimi K3 - early-access MoonshotAI multimodal MoE.
+  // Metadata is sourced from Neuralwatt's authenticated model catalog.
   {
-    id: "deepseek-v4-flash",
-    name: "DeepSeek V4 Flash (Canary)",
+    id: "kimi-k3",
+    name: "Kimi K3",
     reasoning: true,
     input: ["text", "image"],
     cost: {
-      input: 0.14,
-      output: 0.28,
-      cacheRead: 0.0028,
+      input: 3,
+      output: 15,
+      cacheRead: 0.3,
       cacheWrite: 0,
     },
-    contextWindow: 1_000_000,
-    maxTokens: 384_000,
+    contextWindow: 1048560,
+    maxTokens: 65536,
     thinkingLevelMap: {
-      off: "none",
-      minimal: "low",
-      low: "low",
+      minimal: null,
+      low: null,
       medium: "medium",
-      high: "high",
+      high: null,
       xhigh: null,
-      max: "max",
     },
     compat: {
       supportsDeveloperRole: false,

package/extensions/provider/models/legacy.ts

@@ -12,6 +12,7 @@ export const LEGACY_MODEL_ALIAS_MAP = {
   "moonshotai/Kimi-K2.6": "kimi-k2.6",
   "Qwen/Qwen3.5-397B-A17B-FP8": "qwen3.5-397b",
   "Qwen/Qwen3.6-35B-A3B": "qwen3.6-35b",
+  "nvidia/Gemma-4-31B-IT-NVFP4": "gemma-4-31b",
 } as const;
 
 export const LEGACY_NEURALWATT_MODEL_IDS = new Set<string>(

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

@@ -3,23 +3,33 @@ import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
 // Public models returned by https://api.neuralwatt.com/v1/models (unauthenticated view).
 // Pricing, capabilities, and limits are sourced from the API metadata fields.
 export const NEURALWATT_MODELS: ProviderModelConfig[] = [
-  // Gemma 4 31B - NVIDIA NVFP4 checkpoint
+  // DeepSeek V4 Flash - DeepSeek
   {
-    id: "nvidia/Gemma-4-31B-IT-NVFP4",
-    name: "Gemma 4 31B (NVFP4)",
-    reasoning: false,
-    input: ["text", "image"],
+    id: "deepseek-v4-flash",
+    name: "DeepSeek V4 Flash",
+    reasoning: true,
+    input: ["text"],
     cost: {
-      input: 0.144,
-      output: 0.42,
-      cacheRead: 0.036,
+      input: 0.104,
+      output: 0.207,
+      cacheRead: 0.026,
       cacheWrite: 0,
     },
-    contextWindow: 262128,
-    maxTokens: 16384,
+    contextWindow: 1048560,
+    maxTokens: 65536,
+    thinkingLevelMap: {
+      off: "none",
+      minimal: "low",
+      low: "low",
+      medium: "medium",
+      high: "high",
+      xhigh: null,
+      max: "max",
+    },
     compat: {
       supportsDeveloperRole: false,
       maxTokensField: "max_tokens",
+      requiresReasoningContentOnAssistantMessages: true,
     },
   },
   // Gemma 4 31B - Google, served from NVIDIA's NVFP4 checkpoint

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aliou/pi-neuralwatt",
-  "version": "0.10.4",
+  "version": "0.10.5",
   "license": "MIT",
   "type": "module",
   "private": false,