LLMPrivate

NVIDIA Nemotron 3 Ultra

NVIDIA's flagship open-weights frontier model — a 550B-parameter hybrid Mamba-MoE architecture built for agentic reasoning, tool use, and long-context throughput.

Maker
NVIDIA Corporation
Modality
Text
Context
256K tokens
Open weights
Yes — OpenMDW-1.1 license

Overview

What is NVIDIA Nemotron 3 Ultra

NVIDIA Nemotron 3 Ultra is a frontier-scale 550B-parameter (55B active) open-weights language model released in June 2026. It combines a hybrid Mamba-2, MoE, and Attention architecture with Multi-Token Prediction for efficient high-throughput agentic reasoning, native tool use, and long-context analysis.

Running it privately on Venice

On Venice, Nemotron 3 Ultra runs privately with zero retention of your prompts — no surveillance, no training data harvesting. You get native tool use, reasoning, and web search capabilities with the efficiency of an open-weights model, processed on Venice's permissionless infrastructure without a Big Tech account.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • Open weights under OpenMDW-1.1: full checkpoints, training data, and recipes released for self-hosting and fine-tuning.
  • Frontier agentic performance with native tool use, reasoning, and web search support for complex multi-step workflows.
  • Hybrid Mamba-2 + MoE + Attention architecture with Multi-Token Prediction delivers up to ~6× higher inference throughput than comparable open LLMs.
  • Configurable reasoning budget via chat template (enable_thinking), letting you trade speed for depth per task.
  • Strong multilingual coverage across English, French, Spanish, Italian, German, Japanese, Korean, Hindi, Brazilian Portuguese, and Chinese.
  • Designed for up to 1M token context length and strong long-context RULER performance, with 256K available on Venice.
Limitations
  • Not uncensored: it carries safety alignment and will refuse certain requests, unlike fully uncensored models.
  • Massive hardware footprint for self-hosting: BF16 requires 8× GB200/B200 or 16× H100, making local inference expensive.
  • Venice currently offers a 256K context window and fp8 quantization, which is below the native 1M context and BF16 precision.
  • Premium output pricing ($3.13/1M tokens) is higher than efficient rivals like DeepSeek V3.2.
  • As a cutting-edge MoE architecture, it may have less third-party tooling and framework support than standard dense transformers.

Capabilities

What it supports

  • Tool use / function calling
  • Vision (image input)
  • Reasoning
  • Web search
  • Code-optimized
  • Structured output (JSON schema)
  • Audio input
  • Video input
  • Multiple image inputs
  • Log probabilities

Specifications

Datasheet

Maker
NVIDIA Corporation
Released
June 4, 2026
Architecture
LatentMoE — Mamba-2 + MoE + Attention hybrid with Multi-Token Prediction (MTP)
Parameters
550B total / 55B active
Open weights
Yes — OpenMDW-1.1 license
Context window
256K tokens
Max output
32.768K tokens
Capabilities
Function calling, Reasoning, Web search
Privacy on Venice
Private — zero retention
Available on Venice since
Jun 2026

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia-nemotron-3-ultra-550b-a55b",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.63 per 1M input tokens and $3.13 per 1M output tokens.

Input / 1M tokens
$0.63
Per 1M tokens
Output / 1M tokens
$3.13
Per 1M tokens
Cached input / 1M
$0.19
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelPrice (Venice)Context windowOpen weightsBest for
NVIDIA Nemotron 3 Ultra$0.63 in · $3.13 out / 1M256K tokensYesAgentic reasoning & throughput
Kimi K2.6$0.75 in · $3.50 out / 1M256K tokensYesLong-context reasoning & coding
DeepSeek V3.2$0.33 in · $0.48 out / 1M160K tokensYesEfficiency & math/code
Claude Sonnet 4.6$3.60 in · $18 out / 1M1M tokensNoGeneral-purpose closed model

NVIDIA's most capable open-weights model — hybrid Mamba-MoE with configurable reasoning and native tool use.

Use cases

What it is good for

  1. 01Long-running autonomous agents that plan, call tools, and maintain state across many turns.
  2. 02High-stakes RAG over very long documents and legal or scientific corpora.
  3. 03Multilingual reasoning and analysis across code, math, and science benchmarks.
  4. 04Enterprise workflows requiring open weights and data sovereignty for compliance.
  5. 05Code generation and review with extended context for large repositories.

Prompting

Getting better results

Toggle reasoning depth with the chat template (enable_thinking=True/False) depending on whether you need deep analysis or fast answers.

Leverage the 256K context window for large document RAG — feed full papers or codebases in a single prompt.

Use structured output (JSON schema) when building agents that need machine-readable tool arguments.

For multi-turn agent tasks, keep the conversation history concise to maximize available context for reasoning.

FAQ

Frequently asked questions

NVIDIA Nemotron 3 Ultra is a 550B-parameter (55B active) open-weights frontier language model released in June 2026. It uses a hybrid Mamba-2, MoE, and Attention architecture with Multi-Token Prediction for high-throughput agentic reasoning, tool use, and long-context analysis.

Venice charges $0.63 per 1M input tokens and $3.13 per 1M output tokens. Cached input is $0.19 per 1M tokens. There is no subscription required — you pay per token with credits.

You can try it free on Venice — new accounts receive welcome credits and a daily free tier, with no credit card required. Continued use is billed per token in credits.

It is open weights under the OpenMDW-1.1 license. NVIDIA releases the base, post-trained, and quantized checkpoints along with training data and recipes, but it is not an open-source code project in the traditional sense.

Yes. On Venice it supports function calling, reasoning, and web search, plus structured JSON schema output for building agents and automated workflows.

Nemotron 3 Ultra emphasizes agentic throughput and configurable reasoning with a hybrid Mamba-MoE architecture. Kimi K2.6 is also open weights with strong long-context and coding skills. Choose Nemotron for complex multi-step agents and Kimi for deep coding tasks.

Venice currently hosts Nemotron 3 Ultra with a 256K token context window and up to 32,768 tokens max output. NVIDIA's native model supports up to 1M tokens for deployments with sufficient hardware.

Yes. The model supports a configurable reasoning mode via the chat template (enable_thinking=True/False), letting you control the depth of analysis and token usage per request.

No. It is a safety-aligned model and will refuse certain harmful or restricted requests. It is not uncensored, though Venice runs it privately without content monitoring or logging.

Run NVIDIA Nemotron 3 Ultra privately

No prompt logging. No data used for training.