Nemotron 3 Ultra

Nemotron 3 Ultra — Open Frontier-Reasoning LLM (Text Generation)

What is Nemotron 3 Ultra?

Nemotron 3 Ultra is NVIDIA's largest open-weight language model, built for long-running agentic workflows rather than single-turn chat. It is a 550-billion-parameter Mixture-of-Experts model that activates only 55 billion parameters per token, using a hybrid Mamba-Transformer architecture with LatentMoE routing and Multi-Token Prediction. On Segmind it runs as a synchronous text-to-text API: send a prompt, receive the generated text directly. It supports a context window of up to 1 million tokens, making it well suited to reasoning over large codebases, long documents, and multi-source research in a single request.

Key Features

• •550B total / 55B active Mixture-of-Experts for frontier accuracy at high throughput
• •Hybrid Mamba-2 plus Attention layers for efficient long-context handling and precise recall
• •Up to 1M-token context window; leads open LLMs on RULER at 1M context length
• •Controllable reasoning: reasoning on, off, and low-effort modes with an optional reasoning budget
• •Strong tool calling, function calling, and structured multi-step planning for agent harnesses
• •Multilingual output across English plus ten additional languages and 43 programming languages

Best Use Cases

Nemotron 3 Ultra shines as the planner or orchestrator inside an agent system. Reach for it on the genuinely hard calls: sustaining architectural decisions across long coding sessions, synthesizing contradictory evidence across hundreds of research sources, or verifying designs across thousands of constraints. It is a strong fit for coding agents, enterprise document RAG, deep research automation, long-document analysis, and multi-agent orchestration where accuracy and long-context reasoning matter more than raw speed. For simple chat or high-volume routine calls, a smaller model is usually the better choice.

Prompt Tips and Output Quality

Because the model first generates a reasoning trace and then a final answer, give it a clear goal, explicit structure, and any constraints up front. Ask for numbered steps or headed sections when you want a structured report. For high-volume steps, prefer low-effort reasoning to cut tokens; reserve full reasoning for complex planning. If you need strict JSON or schema output, validate and retry inside your harness, as structured-output reliability can vary.

How to Use Nemotron 3 Ultra

Nemotron 3 Ultra is a reasoning-first, agent-oriented model. The Segmind endpoint is synchronous and takes a single prompt field, so effective use comes down to how you frame that prompt and where you deploy the model in your stack.

Write prompts for a reasoning model

The model generates an internal reasoning trace before its final answer, so it rewards clear, well-scoped instructions. State the goal, the required output format, and any constraints explicitly. For analytical work, ask for numbered steps or headed sections to get a structured report. For coding, describe the repository context, the target behavior, and the expected output format. Avoid vague one-line prompts if you want frontier-quality results.

Match the model to the task

Use Nemotron 3 Ultra as the planner or orchestrator on the hardest calls: multi-step planning, synthesizing many sources, reviewing large codebases, and long-document analysis. Route routine, high-volume, or latency-sensitive steps to a smaller model such as Nemotron 3 Super or Nano. Do not send every request to Ultra; that raises cost without improving results. A 1M-token context does not mean every request should fill it, since long inputs still affect latency and reliability.

Get reliable structured output

Structured-output reliability can vary, so if you need strict JSON or a fixed schema, validate the response in your harness and retry on failure. Building a validate-and-retry loop around the model is the recommended pattern for agentic pipelines.

Recommended Settings

• •prompt: Provide full context in one message. For reports, request explicit sections; for code, specify language, files, and output format.
• •Reasoning depth: Prefer lighter reasoning for simple, high-volume steps and full reasoning for complex planning and verification.
• •Context: Include long documents or codebases directly when the task needs cross-document reasoning; otherwise keep inputs tight for speed.
• •Deployment: Best inside an agent harness that can plan, call tools, read observations, and recover from errors across many turns.

Is Nemotron 3 Ultra open weights?

Yes. NVIDIA released the weights, training recipes, and datasets under a permissive open model license.

What context length does it support?

Up to 1 million tokens, with strong performance on long-context retrieval benchmarks.

Is it good for coding agents?

Yes. It is post-trained for agent harnesses and performs well on SWE-bench and Terminal-Bench style tasks with efficient token usage.

How does it compare to other open models?

It is the leading US open-weight model on the Artificial Analysis Intelligence Index, trailing the Chinese-led open frontier such as Kimi K2.6 on raw intelligence while leading on throughput.

Can it call tools?

Yes. It supports streaming reasoning and incremental tool and function calls, designed for multi-turn agent loops.

What are good alternatives?

Nemotron 3 Super or Nano for cheaper, faster steps, and Kimi K2.6 or Qwen 3.5 when maximum benchmark intelligence is the priority.