Skip to main content
  • Pricing
Tutorials

Running Nemotron 3 Super with Ollama and OpenClaw: specs, setup steps and limits

By Oakgen TeamPublished 3 min read

Share
Running Nemotron 3 Super with Ollama and OpenClaw: specs, setup steps and limits

Nemotron 3 Super is an open NVIDIA model with 120 billion total and 12 billion active parameters. Ollama publishes it as nemotron-3-super, and OpenClaw can use any Ollama model through Ollama's native API. As of 2026-11-19, the local download is 87GB, so the practical questions are memory, context size and one configuration rule. This guide sticks to what those vendors document and has not been tested on hardware.

What the model is

NVIDIA's launch post (March 11, 2026) describes Nemotron 3 Super as a hybrid mixture-of-experts model, with only 12 billion of its 120 billion parameters active at inference. The Hugging Face card gives the architecture as LatentMoE, a Mamba-2, MoE and attention hybrid with Multi-Token Prediction, a context length of up to 1 million tokens, support for English, French, German, Italian, Japanese, Spanish and Chinese, and a reasoning mode that can be switched on or off in the chat template. NVIDIA's throughput claims, such as up to 5x higher than the previous Nemotron Super, are vendor figures.

The same card lists a minimum GPU requirement of 2x H100-80GB for the FP8 checkpoint. That is NVIDIA's requirement for its own checkpoint, not for Ollama's quantized build.

What Ollama gives you

Ollama's library page shows ollama run nemotron-3-super, a REST example against port 11434, and capability labels for tools, thinking and cloud. Its tag list shows nemotron-3-super:latest and :120b at 87GB with a 256K context window, and a :cloud tag. Under Applications it lists a launch command for OpenClaw: ollama launch openclaw --model nemotron-3-super.

Two numbers are easy to confuse. NVIDIA's native context is up to 1 million tokens. The Ollama tags list 256K. And OpenClaw, as the next section shows, may cap what it sends well below either figure.

Connecting OpenClaw to Ollama

OpenClaw's Ollama pages set out the rules:

  • Use the native API. Point baseUrl at http://host:11434, not /v1. The docs warn that the OpenAI-compatible URL breaks tool calling.
  • Onboard. Run openclaw onboard, choose Ollama, then Cloud + Local or Local only. For hosted-only access there is a separate Ollama Cloud choice.
  • Know what auto-discovery sees. On a fresh guided setup, discovery considers only models already loaded in memory, with tool support and at least 16K of context confirmed by Ollama's /api/show. An installed but idle model is not an automatic candidate.
  • Verify. openclaw models list --provider ollama, then openclaw models set ollama/<model> or set agents.defaults.model.primary in config.
  • Credentials. Loopback, private-network and .local Ollama URLs need no real token; OpenClaw uses an ollama-local marker. Public hosts and ollama.com need a real key.

Settings that matter for a very large model

ProblemWhat the OpenClaw docs suggest
Slow first response or timeout when the model loadsSet timeoutSeconds (their example uses 300) and a keep_alive such as 15m on the model entry
Out of memory or slow first tokenCap contextTokens and params.num_ctx (their example uses 32768) and lower maxTokens
Tool calls appearing as plain textUse native mode with api: "ollama"; if a small model still fails, set compat.supportsTools: false
Reasoning costThinking levels (/think low to max) are exposed for models whose /api/show reports a thinking capability

The advanced page adds that local discovery normally caps contextTokens at 32,768 or the model's smaller native window, so a model with a larger window will not automatically receive all of it.

Reading the benchmarks fairly

Ollama's page carries a benchmark table in the model's readme text, which compares Nemotron 3 Super with Qwen3.5-122B-A10B and GPT-OSS-120B. The numbers are mixed. Nemotron scores 93.67 on HMMT Feb25 without tools against 91.40 for Qwen and 90.00 for GPT-OSS, and 81.19 on LiveCodeBench v5 against 78.93 for Qwen. Qwen3.5-122B-A10B leads on others, for example Terminal Bench Core 2.0 (37.50 against 31.00) and the TauBench V2 average (74.53 against 61.15). On RULER at 1M tokens Nemotron lists 91.75 against 22.30 for GPT-OSS, but Ollama's local tags stop at 256K. These are results published with the model, not independent tests.

Before you let it act

OpenClaw is an agent that calls tools, and its docs show a non-interactive onboarding flag named --accept-risk. Read what each tool can do on your machine before enabling it, and start with a task whose side effects you can undo. That is general caution, not a statement from the vendors. Whether the setup fits your machine comes down to the 87GB size and the context settings above.

Take this further with Oakgen Agent

Agent can search the web and fetch the vendor docs, summarize requirements and settings, and save a checklist as a markdown or pdf file; it cannot install or run models on your computer.

Coming soon: Pasting a YouTube link and getting a summary of the video inside Agent is not available yet.

Looking for a recap of one specific YouTube video? The YouTube Summarizer does that today.

Open this task in Oakgen Agent

The prompt is filled in for you. Nothing is sent until you press send.

Open in Oakgen Agent

Frequently asked questions

How big is Nemotron 3 Super in Ollama?

As of 2026-11-19, Ollama's library page lists nemotron-3-super:latest and nemotron-3-super:120b at 87GB with a 256K context window, plus a nemotron-3-super:cloud tag. NVIDIA describes the model as 120 billion parameters in total with 12 billion active, and a native context of up to 1 million tokens, so Ollama's local tags expose a smaller window than the model supports natively.

Which Ollama endpoint should OpenClaw use?

OpenClaw's documentation says it talks to Ollama's native API at /api/chat, not the OpenAI-compatible /v1 endpoint. It warns that using http://host:11434/v1 breaks tool calling and can make models print raw tool-call JSON as plain text, and says to set baseUrl to http://host:11434 with no /v1.

What license does Nemotron 3 Super use?

NVIDIA's Hugging Face model card for the FP8 checkpoint says use is governed by the NVIDIA Nemotron Open Model License and describes the model as ready for commercial use. NVIDIA's launch post says the model is released with open weights under a permissive license. Read the license text itself before deploying commercially.

Can I run it without a big local machine?

Ollama's page lists a nemotron-3-super:cloud tag, and OpenClaw's docs describe Cloud + Local and Ollama Cloud setups in which a signed-in Ollama host or ollama.com serves hosted models. Whether a given model is available that way depends on Ollama's current catalog, so check the tag list before relying on it.

Sources

  1. OpenClaw docs, Ollama provider index (read 2026-11-19)
  2. OpenClaw docs, Ollama setup (read 2026-11-19)
  3. OpenClaw docs, Ollama advanced configuration (read 2026-11-19)
  4. OpenClaw docs, Ollama troubleshooting (read 2026-11-19)
  5. Ollama library, nemotron-3-super (read 2026-11-19)
  6. NVIDIA Blog, New NVIDIA Nemotron 3 Super Delivers 5x Higher Throughput for Agentic AI, 2026-03-11 (read 2026-11-19)
  7. Hugging Face, nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8 model card (read 2026-11-19)
Share