Nemotron 3 Ultra for MCP: Specs & Tool Calling — Test Free

What Nemotron 3 Ultra 550B is

Nemotron 3 Ultra is a 550B-parameter mixture-of-experts model with 55B active per token. It mixes Mamba layers, which handle long sequences cheaply, with Transformer layers for precise recall. It uses NVIDIA’s LatentMoE routing and multi-token prediction for faster generation, and it is released in NVFP4 as well as BF16. The weights are open under OpenMDW-1.1, the Linux Foundation’s permissive licence for model distributions. It is text-only. Nemotron 3 Super (120B, 12B active) is the smaller model in the same family, and both are available here.

What NVIDIA says

PinchBench: 91%. Terminal-Bench 2.0: 54%. EnterpriseOps-Gym: 33%.
SWE-bench Verified: between 65% and 70.4%, depending on the harness.
NVIDIA says it cuts the cost of completing SWE-bench and Terminal-Bench 2.0 tasks by up to 30% by using fewer tokens per turn.

These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.

How it reaches MCP

MCP Playground’s Agent Studio: paste your server URL above and run Nemotron 3 Ultra against it in the browser, with no NVIDIA key
NVIDIA’s hosted API at build.nvidia.com, which is OpenAI-compatible, so MCP clients that accept a custom endpoint can point at it with a base-URL change
The open weights, served with vLLM, SGLang or TensorRT-LLM, behind any MCP client

What to watch for

NVIDIA benchmarks the model at 1M tokens, but OpenRouter serves it with a 262K window. If you plan around the 1M figure, check the window of the endpoint you actually call.
OpenRouter lists tools and tool_choice for it but not parallel_tool_calls. A client that fans out several calls in one turn may see them come back one at a time.
Its Terminal-Bench 2.0 score (54%) is well below the closed frontier models. The case for it is open weights and cost per task, so measure tokens used as well as whether the task completed.

Run Nemotron 3 Ultra and Nemotron 3 Super on the same four-step task against the complex-schema mock: create_user_profile, process_order, analyze_data, then configure_workflow. Same family at a tenth of the size, so the gap on your tools tells you whether the Ultra is worth it. All the mock servers →

Available here

Model	Model ID
Nemotron 3 Ultra 550B	`nvidia/nemotron-3-ultra-550b-a55b`

Go deeper

Frequently asked questions

Is Nemotron 3 Ultra open source?

The weights are public on Hugging Face under OpenMDW-1.1, a permissive licence from the Linux Foundation. NVIDIA publishes them in BF16 and NVFP4.

Does Nemotron 3 Ultra support tool calling?

Yes. OpenRouter lists tools and tool_choice, and NVIDIA built it for multi-turn agent work. MCP servers work through any client that turns MCP tools into function calls.

What hardware does Nemotron 3 Ultra need?

NVIDIA says the NVFP4 build runs on Hopper, Blackwell and Ampere GPUs. At 550B parameters it is still a multi-GPU model. To try it against an MCP server without that, use it here.

Nemotron 3 Ultra or Nemotron 3 Super?

Ultra is the larger model with 55B active parameters; Super has 12B active and is free on OpenRouter. Both have 262K windows here. Run the same prompt on each against your server and compare.

Nemotron 3 Ultra 550B