fix(inference): refresh NVIDIA Endpoints featured models

Signed-off-by: Ho Lim <subhoya@gmail.com>

Author: HOYALIM

ci/test-file-size-budget.json

@@ -10,7 +10,7 @@
     "test/install-preflight.test.ts": 4006,
     "test/nemoclaw-start.test.ts": 5043,
     "test/onboard-messaging.test.ts": 2062,
-    "test/onboard-selection.test.ts": 6867,
+    "test/onboard-selection.test.ts": 6864,
     "test/onboard.test.ts": 4774,
     "test/policies.test.ts": 2489
   }

docs/inference/inference-options.mdx

@@ -61,10 +61,12 @@ The first six are always available.
 Ollama appears when you have installed or started it on the host.
 Local vLLM appears when NemoClaw detects a running vLLM server.
 The managed install/start vLLM entry appears by default on DGX Spark and DGX Station, and appears on generic Linux NVIDIA GPU hosts after opt-in.
+During interactive NVIDIA Endpoints onboarding, NemoClaw fetches NVIDIA's featured-models catalog for the model picker.
+If the catalog cannot be fetched or returns no safe model IDs, the wizard uses the curated fallback list shown below.
 
 | Option | Description | Curated models |
 |--------|-------------|----------------|
-| NVIDIA Endpoints | Routes to models hosted on [build.nvidia.com](https://build.nvidia.com). You can also enter any model ID from the catalog. Set `NVIDIA_INFERENCE_API_KEY`. | Nemotron 3 Super 120B, Nemotron 3 Ultra 550B, GLM-5.1, MiniMax M2.7, GPT-OSS 120B, DeepSeek V4 Pro |
+| NVIDIA Endpoints | Routes to models hosted on [build.nvidia.com](https://build.nvidia.com). You can also enter any model ID from the catalog. Set `NVIDIA_INFERENCE_API_KEY`. | Nemotron 3 Ultra 550B, Nemotron 3 Super 120B, GLM 5.1, Kimi K2.6, Minimax M2.7 |
 | OpenAI | Routes to the OpenAI API. Set `OPENAI_API_KEY`. | `gpt-5.4`, `gpt-5.4-mini`, `gpt-5.4-nano`, `gpt-5.4-pro-2026-03-05` |
 | Other OpenAI-compatible endpoint | Routes to any server that implements `/v1/chat/completions`. NemoClaw uses `/v1/chat/completions` at runtime by default; set `NEMOCLAW_PREFERRED_API=openai-responses` to allow `/v1/responses` for proxies that implement it, such as some llama.cpp builds. The wizard prompts for a base URL and model name. The adapter is validated against OpenRouter (refer to the status table above); behavior on other OpenAI-compatible proxies, gateways, and self-hosted implementations such as LocalAI or llama.cpp may vary. When you enable Telegram messaging, onboarding also runs a bounded sandbox-side smoke check through `https://inference.local/v1/chat/completions`. Set `COMPATIBLE_API_KEY`. | You provide the model name. |
 | Anthropic | Routes to the Anthropic Messages API. Set `ANTHROPIC_API_KEY`. | `claude-sonnet-4-6`, `claude-haiku-4-5`, `claude-opus-4-6` |
@@ -78,18 +80,16 @@ The managed install/start vLLM entry appears by default on DGX Spark and DGX Sta
 
 Use this table as starter guidance when selecting a curated cloud model during onboarding.
 The provider catalog remains authoritative for exact context-window limits, availability, and current pricing.
+The table covers the fallback curated choices; the fetched NVIDIA featured-models catalog can include newer featured models.
 The relative labels below are qualitative and compare models within the curated onboarding choices, not across every model a provider offers.
 
 | Model | Best-for task type | Relative latency | Tool-use quality | Context-window fit | Relative cost |
 |---|---|---|---|---|---|
-| `nvidia/nemotron-3-super-120b-a12b` | Default hosted agent work, multi-step planning, and tool-heavy shell workflows | Medium | Strong default for OpenClaw tool loops | Large agent context | Medium |
 | `nvidia/nemotron-3-ultra-550b-a55b` | Quality-sensitive reasoning, careful synthesis, and complex reviews | Higher | Strong for complex tool plans | Large agent context | Higher |
-| `nvidia/nemotron-3-nano-omni-30b-a3b-reasoning` | Reasoning-first and multimodal experiments where a compact hosted model is enough | Medium | Good after the smoke probe confirms final-answer content | Large agent context | Medium |
+| `nvidia/nemotron-3-super-120b-a12b` | Default hosted agent work, multi-step planning, and tool-heavy shell workflows | Medium | Strong default for OpenClaw tool loops | Large agent context | Medium |
 | `z-ai/glm-5.1` | General chat, multilingual text work, and fast iteration | Low-to-medium | Good for straightforward tool loops | Large agent context | Low-to-medium |
-| `minimaxai/minimax-m2.7` | Long-form writing, multi-turn assistant work, and broad instruction following | Medium | Good for structured assistant turns | Large agent context | Medium |
 | `moonshotai/kimi-k2.6` | Coding tasks and shell-heavy agent trajectories | Medium | Strong with NemoClaw's Kimi tool-call compatibility path | Large-context friendly | Medium |
-| `openai/gpt-oss-120b` | Hosted open-weight style experimentation and cost-aware general agents | Medium | Good when provider-side tool calling is enabled | Large agent context | Medium |
-| `deepseek-ai/deepseek-v4-pro` | Code, math, and reasoning-heavy problem solving | Medium-to-high | Strong when the endpoint supports tool calls | Large agent context | Medium-to-high |
+| `minimaxai/minimax-m2.7` | Long-form writing, multi-turn assistant work, and broad instruction following | Medium | Good for structured assistant turns | Large agent context | Medium |
 | `gpt-5.4` | Default OpenAI-backed agent work and general high-quality reasoning | Medium | Strong | Large agent context | Medium-to-high |
 | `gpt-5.4-mini` | Latency-sensitive routine automation and repeated helper calls | Low | Good | Medium-to-large context | Low |
 | `gpt-5.4-nano` | Very low-latency classification, routing, extraction, and small helper tasks | Very low | Basic to good for simple tool loops | Medium context | Very low |

src/lib/inference/config.test.ts

@@ -6,6 +6,7 @@ import { describe, expect, it } from "vitest";
 // Import source directly so tests cannot pass against a stale build.
 import {
   CLOUD_MODEL_OPTIONS,
+  DEFAULT_CLOUD_MODEL,
   DEFAULT_HERMES_PROVIDER_MODEL,
   DEFAULT_OLLAMA_MODEL,
   DEFAULT_ROUTE_CREDENTIAL_ENV,
@@ -25,16 +26,29 @@ import {
 
 describe("inference selection config", () => {
   it("exposes the curated cloud model picker options", () => {
-    expect(CLOUD_MODEL_OPTIONS.map((option: { id: string }) => option.id)).toEqual([
-      "nvidia/nemotron-3-super-120b-a12b",
-      "nvidia/nemotron-3-ultra-550b-a55b",
+    expect(CLOUD_MODEL_OPTIONS).toEqual([
+      { id: "nvidia/nemotron-3-ultra-550b-a55b", label: "Nemotron 3 Ultra 550B" },
+      { id: "nvidia/nemotron-3-super-120b-a12b", label: "Nemotron 3 Super 120B" },
+      { id: "z-ai/glm-5.1", label: "GLM 5.1" },
+      { id: "moonshotai/kimi-k2.6", label: "Kimi K2.6" },
+      { id: "minimaxai/minimax-m2.7", label: "Minimax M2.7" },
+    ]);
+    expect(CLOUD_MODEL_OPTIONS.map((option: { id: string }) => option.id)).not.toContain(
       "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
-      "z-ai/glm-5.1",
-      "minimaxai/minimax-m2.7",
-      "moonshotai/kimi-k2.6",
+    );
+    expect(CLOUD_MODEL_OPTIONS.map((option: { id: string }) => option.id)).not.toContain(
       "openai/gpt-oss-120b",
+    );
+    expect(CLOUD_MODEL_OPTIONS.map((option: { id: string }) => option.id)).not.toContain(
       "deepseek-ai/deepseek-v4-pro",
-    ]);
+    );
+  });
+
+  it("keeps the NVIDIA Endpoints default on Nemotron 3 Super", () => {
+    expect(DEFAULT_CLOUD_MODEL).toBe("nvidia/nemotron-3-super-120b-a12b");
+    expect(CLOUD_MODEL_OPTIONS.map((option: { id: string }) => option.id)).toContain(
+      "nvidia/nemotron-3-super-120b-a12b",
+    );
   });
 
   it("aligns Hermes Provider defaults with the Hermes Agent Nous catalog", () => {

src/lib/inference/config.ts

@@ -53,14 +53,11 @@ export const HERMES_PROVIDER_MODEL_OPTIONS = [
 ] as const;
 export const DEFAULT_HERMES_PROVIDER_MODEL = HERMES_PROVIDER_MODEL_OPTIONS[0];
 export const CLOUD_MODEL_OPTIONS = [
-  { id: "nvidia/nemotron-3-super-120b-a12b", label: "Nemotron 3 Super 120B" },
   { id: "nvidia/nemotron-3-ultra-550b-a55b", label: "Nemotron 3 Ultra 550B" },
-  { id: "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning", label: "Nemotron 3 Nano Omni 30B" },
-  { id: "z-ai/glm-5.1", label: "GLM-5" },
-  { id: "minimaxai/minimax-m2.7", label: "MiniMax M2.7" },
+  { id: "nvidia/nemotron-3-super-120b-a12b", label: "Nemotron 3 Super 120B" },
+  { id: "z-ai/glm-5.1", label: "GLM 5.1" },
   { id: "moonshotai/kimi-k2.6", label: "Kimi K2.6" },
-  { id: "openai/gpt-oss-120b", label: "GPT-OSS 120B" },
-  { id: "deepseek-ai/deepseek-v4-pro", label: "DeepSeek V4 Pro" },
+  { id: "minimaxai/minimax-m2.7", label: "Minimax M2.7" },
 ];
 export const DEFAULT_ROUTE_PROFILE = "inference-local";
 export const DEFAULT_ROUTE_CREDENTIAL_ENV = "OPENAI_API_KEY";

src/lib/inference/model-prompts.test.ts

@@ -33,14 +33,14 @@ describe("model prompt helpers", () => {
     expect(result).toBe("llama");
   });
 
-  it("returns DeepSeek V4 Pro from the default cloud model menu", async () => {
-    const promptFn = promptSequence(["8"]);
+  it("returns Minimax M2.7 from the default cloud model menu", async () => {
+    const promptFn = promptSequence(["5"]);
     const result = await promptCloudModel({
       promptFn,
       writeLine: vi.fn(),
     });
 
-    expect(result).toBe("deepseek-ai/deepseek-v4-pro");
+    expect(result).toBe("minimaxai/minimax-m2.7");
   });
 
   it("validates manual cloud model ids against the saved NVIDIA key", async () => {

src/lib/inference/provider-models.test.ts

@@ -6,14 +6,110 @@ import { describe, expect, it } from "vitest";
 import {
   BUILD_ENDPOINT_URL,
   fetchAnthropicModels,
+  fetchNvidiaFeaturedModels,
   fetchNvidiaEndpointModels,
   fetchOpenAiLikeModels,
+  getNvidiaFeaturedModelPromptOptions,
+  getNvidiaFeaturedModelOptions,
+  NVIDIA_FEATURED_MODELS_URL,
+  parseNvidiaFeaturedModels,
   validateAnthropicModel,
   validateNvidiaEndpointModel,
   validateOpenAiLikeModel,
 } from "./provider-models";
 
 describe("provider model helpers", () => {
+  it("parses NVIDIA featured models from the build catalog snapshot shape", () => {
+    expect(
+      parseNvidiaFeaturedModels(
+        JSON.stringify({
+          "featured-models": [
+            {
+              model: "nvidia/nemotron-3-ultra-550b-a55b",
+              "model-name": "Nemotron 3 Ultra 550B",
+            },
+            {
+              model: "nemotron-3-super-120b-a12b",
+              "model-name": "Nemotron 3 Super 120B",
+            },
+            { model: "z-ai/glm-5.1", "model-name": "GLM 5.1" },
+          ],
+        }),
+      ),
+    ).toEqual([
+      { id: "nvidia/nemotron-3-ultra-550b-a55b", label: "Nemotron 3 Ultra 550B" },
+      { id: "nvidia/nemotron-3-super-120b-a12b", label: "Nemotron 3 Super 120B" },
+      { id: "z-ai/glm-5.1", label: "GLM 5.1" },
+    ]);
+  });
+
+  it("fetches NVIDIA featured models without requiring an API key", () => {
+    const result = fetchNvidiaFeaturedModels({
+      runCurlProbeImpl: (argv) => {
+        expect(argv.at(-1)).toBe(NVIDIA_FEATURED_MODELS_URL);
+        expect(argv.join(" ")).not.toContain("Authorization: Bearer");
+        return {
+          ok: true,
+          httpStatus: 200,
+          curlStatus: 0,
+          body: JSON.stringify({
+            "featured-models": [{ model: "moonshotai/kimi-k2.6", "model-name": "Kimi K2.6" }],
+          }),
+          stderr: "",
+          message: "",
+        };
+      },
+    });
+
+    expect(result).toEqual({
+      ok: true,
+      models: [{ id: "moonshotai/kimi-k2.6", label: "Kimi K2.6" }],
+    });
+  });
+
+  it("falls back to the curated NVIDIA featured model snapshot when the catalog is unavailable", () => {
+    const models = getNvidiaFeaturedModelOptions({
+      runCurlProbeImpl: () => ({
+        ok: false,
+        httpStatus: 503,
+        curlStatus: 0,
+        body: "",
+        stderr: "",
+        message: "HTTP 503",
+      }),
+    });
+
+    expect(models.map((model) => model.id)).toEqual([
+      "nvidia/nemotron-3-ultra-550b-a55b",
+      "nvidia/nemotron-3-super-120b-a12b",
+      "z-ai/glm-5.1",
+      "moonshotai/kimi-k2.6",
+      "minimaxai/minimax-m2.7",
+    ]);
+  });
+
+  it("builds NVIDIA featured model prompt options with the configured default model", () => {
+    const options = getNvidiaFeaturedModelPromptOptions(null, {
+      runCurlProbeImpl: () => ({
+        ok: true,
+        httpStatus: 200,
+        curlStatus: 0,
+        body: JSON.stringify({
+          "featured-models": [
+            { model: "nvidia/nemotron-3-ultra-550b-a55b", "model-name": "Nemotron Ultra" },
+          ],
+        }),
+        stderr: "",
+        message: "",
+      }),
+    });
+
+    expect(options).toEqual({
+      defaultModelId: "nvidia/nemotron-3-super-120b-a12b",
+      cloudModelOptions: [{ id: "nvidia/nemotron-3-ultra-550b-a55b", label: "Nemotron Ultra" }],
+    });
+  });
+
   it("fetches NVIDIA endpoint model ids", () => {
     const result = fetchNvidiaEndpointModels("nvapi-x", {
       runCurlProbeImpl: (argv) => {

src/lib/inference/provider-models.ts

@@ -1,18 +1,23 @@
 // SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 
+import { CLOUD_MODEL_OPTIONS, DEFAULT_CLOUD_MODEL } from "./config";
 import type { CurlProbeResult } from "../adapters/http/probe";
 import { getCurlTimingArgs, runCurlProbe } from "../adapters/http/probe";
 import type { ModelCatalogFetchResult, ModelValidationResult } from "../onboard/types";
+import { isSafeModelId } from "../validation";
 
 // credentials.ts still uses CommonJS-style exports.
 const { normalizeCredentialValue } = require("../credentials/store");
 
 export const BUILD_ENDPOINT_URL = "https://integrate.api.nvidia.com/v1";
+export const NVIDIA_FEATURED_MODELS_URL =
+  "https://assets.ngc.nvidia.com/products/api-catalog/featured-models.json";
 
 export interface ProviderModelOptions {
   runCurlProbeImpl?: (argv: string[]) => CurlProbeResult;
   buildEndpointUrl?: string;
+  featuredModelsUrl?: string;
   /** When "query-param", send the API key as a ?key= URL parameter instead of
    *  an Authorization: Bearer header. Required for Google Gemini which rejects
    *  requests carrying both auth methods. See issue #1960. */
@@ -28,6 +33,32 @@ type ModelCatalogResponse = {
   data?: Array<ModelCatalogItem | null>;
 };
 
+type FeaturedModelCatalogItem = {
+  model?: string | null;
+  "model-name"?: string | null;
+};
+
+type FeaturedModelCatalogResponse = {
+  "featured-models"?: Array<FeaturedModelCatalogItem | null>;
+};
+
+export type FeaturedModelOption = {
+  id: string;
+  label: string;
+};
+
+export type FeaturedModelFetchResult =
+  | {
+      ok: true;
+      models: FeaturedModelOption[];
+    }
+  | {
+      ok: false;
+      message: string;
+      httpStatus: number;
+      curlStatus: number;
+    };
+
 function parseJson<T>(text: string): T {
   return JSON.parse(text);
 }
@@ -51,6 +82,98 @@ function parseModelIds(body: string, itemKeys: Array<keyof ModelCatalogItem> = [
     .filter((value): value is string => Boolean(value));
 }
 
+/**
+ * Normalizes the NVIDIA featured-models catalog IDs into endpoint model IDs.
+ */
+function normalizeFeaturedModelId(model: string): string {
+  const trimmed = model.trim();
+  if (/^nemotron-3-/i.test(trimmed)) {
+    return `nvidia/${trimmed}`;
+  }
+  return trimmed;
+}
+
+/**
+ * Parses NVIDIA's featured-models catalog into safe onboarding menu options.
+ */
+export function parseNvidiaFeaturedModels(body: string): FeaturedModelOption[] {
+  const parsed = parseJson<FeaturedModelCatalogResponse>(body);
+  const featuredModels = parsed["featured-models"];
+  if (!Array.isArray(featuredModels)) {
+    throw new Error('Unexpected featured model catalog response: expected "featured-models" array');
+  }
+
+  return featuredModels
+    .map((item) => {
+      const id = typeof item?.model === "string" ? normalizeFeaturedModelId(item.model) : "";
+      const label = typeof item?.["model-name"] === "string" ? item["model-name"].trim() : "";
+      return id && label && isSafeModelId(id) ? { id, label } : null;
+    })
+    .filter((value): value is FeaturedModelOption => value !== null);
+}
+
+/**
+ * Fetches NVIDIA's public featured-models catalog without requiring credentials.
+ */
+export function fetchNvidiaFeaturedModels(
+  options: ProviderModelOptions = {},
+): FeaturedModelFetchResult {
+  const runCurlProbeImpl = options.runCurlProbeImpl ?? runCurlProbe;
+  const sourceUrl = options.featuredModelsUrl ?? NVIDIA_FEATURED_MODELS_URL;
+  try {
+    const result = runCurlProbeImpl([
+      "-sS",
+      "--connect-timeout",
+      "5",
+      "--max-time",
+      "15",
+      sourceUrl,
+    ]);
+    if (!result.ok) {
+      return {
+        ok: false,
+        message: result.message,
+        httpStatus: result.httpStatus,
+        curlStatus: result.curlStatus,
+      };
+    }
+    return { ok: true, models: parseNvidiaFeaturedModels(result.body) };
+  } catch (error) {
+    return {
+      ok: false,
+      httpStatus: 0,
+      curlStatus: 0,
+      message: error instanceof Error ? error.message : String(error),
+    };
+  }
+}
+
+/**
+ * Returns live featured NVIDIA models, falling back to the curated snapshot.
+ */
+export function getNvidiaFeaturedModelOptions(
+  options: ProviderModelOptions = {},
+): FeaturedModelOption[] {
+  const result = fetchNvidiaFeaturedModels(options);
+  return result.ok && result.models.length > 0 ? result.models : CLOUD_MODEL_OPTIONS;
+}
+
+/**
+ * Builds NVIDIA Endpoints prompt options from the featured-models catalog.
+ */
+export function getNvidiaFeaturedModelPromptOptions(
+  defaultModelId?: string | null,
+  options: ProviderModelOptions = {},
+): {
+  defaultModelId: string;
+  cloudModelOptions: FeaturedModelOption[];
+} {
+  return {
+    defaultModelId: defaultModelId || DEFAULT_CLOUD_MODEL,
+    cloudModelOptions: getNvidiaFeaturedModelOptions(options),
+  };
+}
+
 function toModelCatalogFetchResult(
   result: CurlProbeResult,
   itemKeys: Array<keyof ModelCatalogItem> = ["id"],