Skip to content
Python

NVIDIA NIM: current Nemotron ids with windows, self-hosted Test lists models

Rohit C Prasadaf45cbe49292
diff --git a/coworker/providers/matrix.py b/coworker/providers/matrix.py
index d1a7d767..1fde8e68 100644
--- a/coworker/providers/matrix.py
+++ b/coworker/providers/matrix.py
@@ -335,22 +335,21 @@ MATRIX: dict[str, ModelEntry] = {
"orcarouter:deepseek/deepseek-v4-pro": ModelEntry(
"DeepSeek V4 Pro · via OrcaRouter", _AGENTIC, 1_048_576
),
- # NVIDIA NIM — ids verified against integrate.api.nvidia.com/v1/models (2026-07-25).
- # Tool calling was probed live per model the same day: each returned two well-formed
- # calls in a single assistant turn, so _AGENTIC (incl. parallel) is observed, not
- # assumed. Vision left off — not tested. NVIDIA also fronts kimi-k2.6, but it 404s
- # ("Not found for account") on a standard build.nvidia.com key, so it stays out until
- # that's verifiable.
- "nvidia:nvidia/llama-3.3-nemotron-super-49b-v1.5": ModelEntry(
- "Llama 3.3 Nemotron Super · via NVIDIA"
+ # NVIDIA NIM — NVIDIA's own `vendor/model` namespace, so its models repeat the word in
+ # the routed id. Ids from integrate.api.nvidia.com/v1/models (keyless, read 2026-10-25;
+ # the earlier llama-3.3-nemotron-super-49b, z-ai/glm-5.2 and deepseek-v4-pro ids are
+ # no longer listed). Windows from the build.nvidia.com model pages the same day, which
+ # list "1M" context for all three. Text-only; tool calling as on the OpenRouter and
+ # Bedrock rows for the same models.
+ "nvidia:nvidia/nemotron-3-super-120b-a12b": ModelEntry(
+ "Nemotron 3 Super · via NVIDIA", _AGENTIC, 1_000_000
),
"nvidia:nvidia/nemotron-3-ultra-550b-a55b": ModelEntry(
- "Nemotron 3 Ultra · via NVIDIA"
+ "Nemotron 3 Ultra · via NVIDIA", _AGENTIC, 1_000_000
+ ),
+ "nvidia:nvidia/nemotron-3.5-lightning-30b-a3b": ModelEntry(
+ "Nemotron 3.5 Lightning · via NVIDIA", _AGENTIC, 1_000_000
),
- "nvidia:z-ai/glm-5.2": ModelEntry("GLM-5.2 · via NVIDIA"),
- # ~150s to first response from cold on this route — usable, but noticeably slower
- # than the same model via Together/Fireworks.
- "nvidia:deepseek-ai/deepseek-v4-pro": ModelEntry("DeepSeek V4 Pro · via NVIDIA"),
# -- cloud accounts (models running in the user's own AWS/GCP) ----------------
# Bedrock ids carry a family segment (claude/ → native Anthropic path, other/ →
# Converse) plus AWS's own `-v<n>:<m>` version suffix. Some regions require the
diff --git a/coworker/providers/registry.py b/coworker/providers/registry.py
index 1d582006..a7096d49 100644
--- a/coworker/providers/registry.py
+++ b/coworker/providers/registry.py
@@ -751,7 +751,7 @@ DESCRIPTORS: list[ProviderDescriptor] = [
"nvidia",
"NVIDIA NIM",
base_url="https://integrate.api.nvidia.com/v1",
- recommended_model="nvidia/llama-3.3-nemotron-super-49b-v1.5",
+ recommended_model="nvidia/nemotron-3-super-120b-a12b",
env_key="NVIDIA_API_KEY",
endpoint_help="Prefilled with NVIDIA's hosted NIM endpoint. Point it at a self-hosted NIM container (e.g. http://localhost:8000/v1) to use your own deployment.",
),
@@ -1160,21 +1160,30 @@ def verify_provider_key(
# one, so the usual list-models probe can't tell a good key from a typo. Auth
# is enforced only on inference — and only once the model id routes (a bad
# model 404s before the key is looked at) — so we spend a single token on the
- # recommended model instead.
+ # recommended model instead. A self-hosted NIM (any other endpoint) serves one
+ # model of its own choosing, so the recommended id would 404 there; its
+ # /v1/models is behind the key, so the plain listing is the right probe.
default_base = next(
(f.default for f in d.fields if f.key == "base_url" and f.default), ""
)
base = (base_url or "").strip().rstrip("/") or default_base.rstrip("/")
- resp = httpx.post(
- base + "/chat/completions",
- headers={"Authorization": f"Bearer {key}"},
- json={
- "model": d.recommended_model,
- "messages": [{"role": "user", "content": "hi"}],
- "max_tokens": 1,
- },
- timeout=timeout,
- )
+ if base != default_base.rstrip("/"):
+ resp = httpx.get(
+ base + "/models",
+ headers={"Authorization": f"Bearer {key}"},
+ timeout=timeout,
+ )
+ else:
+ resp = httpx.post(
+ base + "/chat/completions",
+ headers={"Authorization": f"Bearer {key}"},
+ json={
+ "model": d.recommended_model,
+ "messages": [{"role": "user", "content": "hi"}],
+ "max_tokens": 1,
+ },
+ timeout=timeout,
+ )
elif name in ("ark", "ark-agent-plan-cn"):
default_base = next(
(f.default for f in d.fields if f.key == "base_url" and f.default), ""
diff --git a/coworker/server/manager.py b/coworker/server/manager.py
index a3e4e0d7..f16b6250 100644
--- a/coworker/server/manager.py
+++ b/coworker/server/manager.py
@@ -4272,9 +4272,10 @@ class SessionManager:
"qwen": ["qwen3-max", "qwen3-coder-plus", "qwen-plus"],
"xai": ["grok-4.3", "grok-4"],
"mistral": ["mistral-large-latest", "mistral-small-latest"],
- # NIM ids carry their own `vendor/model` namespace; both probed 2026-07-25.
+ # NIM ids carry their own `vendor/model` namespace (integrate.api.nvidia.com/v1/models,
+ # read 2026-10-25).
"nvidia": [
- "nvidia/llama-3.3-nemotron-super-49b-v1.5",
+ "nvidia/nemotron-3-super-120b-a12b",
"nvidia/nemotron-3-ultra-550b-a55b",
],
}
diff --git a/surfaces/gui/src/providers/logos.ts b/surfaces/gui/src/providers/logos.ts
index 3d9cf71f..972020a3 100644
--- a/surfaces/gui/src/providers/logos.ts
+++ b/surfaces/gui/src/providers/logos.ts
@@ -79,7 +79,6 @@ export const PROVIDER_ORDER = [
"openrouter-account",
"fireworks",
"together",
- "nvidia",
"zai",
"kimi",
"deepseek",
@@ -89,6 +88,7 @@ export const PROVIDER_ORDER = [
"xai",
"nebius",
"orcarouter",
+ "nvidia",
];
export function providerRank(name: string): number {
diff --git a/tests/test_provider_verify.py b/tests/test_provider_verify.py
index d1984734..0b47456c 100644
--- a/tests/test_provider_verify.py
+++ b/tests/test_provider_verify.py
@@ -226,15 +226,30 @@ def test_verify_ark_profile_endpoint_override(monkeypatch):
assert cap["url"] == "https://gateway.example/ark/v3/responses"
-def test_verify_nvidia_custom_endpoint(monkeypatch):
- """The descriptor's help text promises a self-hosted NIM container works — only true
- if the override actually reaches the request."""
+def test_verify_nvidia_self_hosted_lists_models(monkeypatch):
+ """The descriptor's help text promises a self-hosted NIM container works. Such a
+ container serves one model of its own, so the hosted probe's recommended id would
+ 404 there; with any endpoint other than the default, Test lists models instead
+ (behind the key on a self-hosted NIM) and never POSTs a completion."""
cap: dict = {}
- _patch_post_refusing_get(monkeypatch, status=200, capture=cap)
- verify_provider_key(
+ _patch_get(monkeypatch, status=200, capture=cap)
+
+ def refuse_post(*_a, **_kw):
+ raise AssertionError("a self-hosted NIM is probed with GET /models, not a completion")
+
+ monkeypatch.setattr("httpx.post", refuse_post)
+ assert verify_provider_key(
"nvidia", api_key="nvapi-x", base_url="http://localhost:8000/v1/"
- )
- assert cap["url"] == "http://localhost:8000/v1/chat/completions"
+ ) == {"ok": True}
+ assert cap["url"] == "http://localhost:8000/v1/models"
+ assert cap["headers"]["Authorization"] == "Bearer nvapi-x"
+
+
+def test_verify_nvidia_self_hosted_rejected_key_is_invalid(monkeypatch):
+ _patch_get(monkeypatch, status=401)
+ assert verify_provider_key(
+ "nvidia", api_key="nvapi-bad", base_url="http://nim.internal:8000/v1"
+ ) == {"ok": False, "error": "Invalid API key."}
def test_verify_network_error_is_clean(monkeypatch):
diff --git a/tests/test_providers.py b/tests/test_providers.py
index 6bd4b305..ea0e732d 100644
--- a/tests/test_providers.py
+++ b/tests/test_providers.py
@@ -505,7 +505,7 @@ def test_matrix_answers_capabilities_for_reseller_ids():
"orcarouter:anthropic/claude-sonnet-5",
# NIM ids are `vendor/model`, so NVIDIA's own models repeat the word in the
# routed id. Deliberate: the `z-ai/` and `deepseek-ai/` rows in matrix.py don't.
- "nvidia:nvidia/llama-3.3-nemotron-super-49b-v1.5",
+ "nvidia:nvidia/nemotron-3-super-120b-a12b",
):
caps = capabilities_for(mid)
assert caps.tools and caps.parallel_tool_calls and caps.streaming