NVIDIA released Nemotron 3 Ultra 550B on June 18, 2026, and the open-weight model went live on Hugging Face the same day. The release includes 550 billion total parameters with 32 billion active parameters per token. The model supports a 128K context window and ships under the NVIDIA Open Model License. NVIDIA describes the license as open-weight but not fully open-source. The weights, tokenizer, and evaluation scripts are public. Training data and the full training code are not public. The model targets reasoning, coding, and tool use workloads. This release follows the company’s Cosmos and Nemotron line.

The vendor announcement is the primary source for this release. NVIDIA did not name a specific GitHub repository with full training code. Instead, the company pointed users to its official NVIDIA homepage and to Hugging Face for model weights. The model card on Hugging Face lists the parameter count, context length, checkpoint formats, and license. Community projects have already published conversion scripts for vLLM and SGLang. The release is part of NVIDIA’s broader push into open-weight foundation models for enterprise and developer use. It gives a US-based answer to open-weight releases from Meta, DeepSeek, and Qwen.

Why it matters is benchmark parity with closed models. NVIDIA reports that Nemotron 3 Ultra 550B scores 85.7 on MMLU Pro, 76.3 on GPQA Diamond, 94.2 on HumanEval, and 93.5 on MATH-500. Those numbers place it within striking distance of Claude Sonnet 4.5 and Gemini 3 Flash on reasoning tasks. At the same time, the license allows commercial use, fine-tuning, and local deployment. That combination removes a major barrier for regulated industries. It also breaks the assumption that top-tier US reasoning models must be accessed through paid APIs. Developers can run the model on their own hardware without monthly token bills, as covered in our guide to self-hosted free AI models.

The release date lands in the middle of a crowded open-weight cycle. Meta’s Llama 4 models and DeepSeek’s V4 both shipped earlier in 2026. Qwen 3.6 and Mistral’s smaller open models are also competing for local inference workloads. Nemotron 3 Ultra stands out because it is the largest US open-weight model from a major infrastructure vendor. It is not a lightweight edge model. It is a data-center class model that requires serious GPU memory. That shifts the frame from free API tier to self-hosted frontier model. For teams already using free AI models or moving away from subscription billing, this is a concrete option.

How Do the Top Options Compare?

Model Parameters Context License Best For
NVIDIA Nemotron 3 Ultra 550B 550B total, 32B active 128K NVIDIA Open Model License US open-weight reasoning and coding
Llama 4 Maverick 400B total, 17B active 1M Meta Llama 4 Community License Multimodal and long context
DeepSeek V4 1.6T total, 32B active 128K MIT Math and reasoning
Qwen 3.6 1.2T total, 32B active 256K Apache 2.0 Coding and agentic workflows

Benchmark and hardware details are based on vendor disclosures and community testing as of June 2026. Active parameter counts reflect tokens processed per forward pass; total parameter counts are larger. Actual VRAM use depends on precision, batch size, and inference framework.

1. NVIDIA Nemotron 3 Ultra 550B , Best US open-weight model for reasoning, coding, and commercial deployment

The model is a 550B parameter mixture of experts architecture. Only 32B parameters are active per forward pass. That keeps inference cost lower than a dense 550B model. The context window is 128K tokens. NVIDIA trained it for tool use, code generation, math, and instruction following. It also supports structured outputs and function calling. The official checkpoints include BF16 and FP8 precision. AWQ and GPTQ quantized versions are available for lower VRAM setups. The model card reports strong performance on long context retrieval tasks. NVIDIA positions this release as part of its agent framework work, detailed in our NVIDIA Nemoclaw open-source agent coverage.

NVIDIA published benchmark results using its own eval harness. On MMLU Pro, Nemotron 3 Ultra 550B scores 85.7. On GPQA Diamond it scores 76.3. HumanEval comes in at 94.2 and MATH-500 at 93.5. These numbers beat Llama 4 Maverick on reasoning and come close to Claude Sonnet 4.5. The model is not the best on every task. It trails specialized coding models on SWE-bench Verified. But it is the strongest US open-weight generalist in June 2026. Independent community testing largely confirms NVIDIA’s numbers in FP8, with small differences in long context recall.

The license is the NVIDIA Open Model License. It allows commercial use, modification, and distribution of derivative models. It does not require your fine-tune to be open-weighted. But it includes export control and use restrictions for defense, surveillance, and certain high-risk applications. That is less permissive than Apache 2.0 but more permissive than many research-only licenses. To run the full BF16 model, you need roughly 1.1TB of VRAM. That means 8x H100 80GB or 4x H200 141GB GPUs. FP8 cuts memory to about 550GB. AWQ 4-bit can fit on 2x A100 80GB with 32K context. You can use vLLM, SGLang, or TensorRT-LLM.

Key strengths:

  • ✅ Strong benchmark parity with closed frontier models on reasoning and math
  • ✅ Commercial-use license allows fine-tuning and private deployment
  • ✅ 128K context handles long documents, code bases, and multi-step agent tasks
  • ✅ FP8 and AWQ checkpoints reduce VRAM requirements for smaller GPU clusters
  • ✅ US-built model with enterprise support path from NVIDIA
  • ❌ 550B total parameters require multi-GPU data center hardware for full precision
  • ❌ Training data and code are not public, so it is open-weight not fully open-source
  • ❌ License includes defense and surveillance use restrictions that may block some users

Who it’s for: Developers and enterprises that need a powerful US open-weight model for local fine-tuning, reasoning, and coding without per-token API billing.

2. Llama 4 Maverick , Best open-weight multimodal model with 1M context

Meta released Llama 4 Maverick earlier in 2026. It has 400 billion total parameters and 17 billion active parameters per token. The context window is 1 million tokens. The model is multimodal, meaning it can process images and text together. The Meta Llama 4 Community License governs use. It allows commercial use but imposes a 700 million monthly active user threshold and acceptable use restrictions. This makes it open-weight, not fully open-source. Details are in our Llama 4 Scout and Maverick release story.

NVIDIA’s Nemotron 3 Ultra beats Llama 4 Maverick on reasoning and math benchmarks. On MMLU Pro, Maverick scores about 82.1 compared with Nemotron’s 85.7. On GPQA Diamond, Maverick scores around 66.5 compared with 76.3. But Maverick wins on multimodal understanding and long context retrieval. Its 1M token context is far larger than Nemotron’s 128K. For document archives, video analysis, and cross-modal search, Maverick remains strong. Meta also offers deep integration with its AI ecosystem and a large community of fine-tuned variants hosted on Hugging Face.

Hardware requirements depend on precision. The full BF16 model needs large GPU memory, similar to Nemotron. Quantized versions can run on fewer GPUs with shorter context. Meta’s license is more restrictive than the NVIDIA Open Model License in some ways. The 700 million user cap can be an issue for very large platforms. On the other hand, Meta’s license does not specifically block defense or surveillance use the way NVIDIA’s license does. Your choice depends on your use case and legal review.

Key strengths:

  • ✅ Massive 1M token context for long document and video tasks
  • ✅ Strong multimodal vision-language capabilities
  • ✅ Meta license allows commercial use for most companies under 700 million monthly active users
  • ✅ Lighter 17B active parameter routing keeps single-token latency low
  • ❌ Lower reasoning and math scores than Nemotron 3 Ultra in independent tests
  • ❌ Meta Llama 4 Community License imposes an active user threshold and policy restrictions
  • ❌ Full model still requires large GPU memory for long context inference

Who it’s for: Teams that need multimodal open-weight capability, very long context, or a Meta ecosystem model.

3. DeepSeek V4 , Best open-weight reasoning and math model with MIT license

DeepSeek V4 released in 2026 as a 1.6T parameter mixture of experts model. It uses 32 billion active parameters per token and supports a 128K context window. The big advantage is the MIT license. That is the most permissive license among top open-weight models. It allows almost any commercial use, modification, and redistribution. DeepSeek is not a US company. That creates export and compliance questions for some enterprises, especially in regulated US government or defense sectors. Our DeepSeek V4 open-source story has more details.

DeepSeek V4 often outscores Nemotron 3 Ultra on math and formal reasoning. On MATH-500, DeepSeek V4 scores around 95.1 compared with Nemotron’s 93.5. On AIME 2026 problems, the gap is wider. That makes DeepSeek V4 the preferred choice for math-heavy workloads, theorem proving, and competitive programming. But Nemotron 3 Ultra is stronger on long-horizon tool use and instruction following in independent tests. DeepSeek’s tool calling can be less consistent when tasks require many sequential API calls. The model is also not tied to NVIDIA’s enterprise support network, though community tooling is robust.

Hardware needs are similar. The full model is very large. But DeepSeek V4 benefits from a huge community that produces aggressive quantization, speculative decoding, and local inference optimizations. If your priority is a permissive license and math reasoning, DeepSeek V4 is hard to beat. If you need a US-based vendor with enterprise support and balanced generalist performance, Nemotron 3 Ultra is the safer pick.

Key strengths:

  • ✅ MIT license is the most permissive among top open-weight models
  • ✅ Class-leading math and formal reasoning benchmark scores
  • ✅ Strong community support for quantization and local inference
  • ✅ Active parameter routing keeps generation fast
  • ❌ Non-US origin triggers export and compliance review for some enterprise deployments
  • ❌ Documentation and model card have less enterprise support than US vendors
  • ❌ Tool calling and function following can be less consistent than Nemotron

Who it’s for: Developers and researchers who need a fully permissive MIT reasoning model and do not have US-only vendor requirements.

4. Qwen 3.6 , Best Apache 2.0 open-weight model for coding and agentic workflows

Qwen 3.6 is an Apache 2.0 licensed open-weight model with 1.2 trillion total parameters and 32 billion active parameters per token. It supports a 256K context window. That is double Nemotron 3 Ultra’s context. Qwen 3.6 is especially strong for coding and agentic tool use. Its HumanEval score is around 93.0, close to Nemotron’s 94.2. But Qwen 3.6 often scores higher on SWE-bench Verified and repository-level coding tasks. Full coverage is in our Qwen 3.6 Apache open-source coding article.

Apache 2.0 is a straightforward commercially friendly license. It does not have the defense and surveillance restrictions found in the NVIDIA Open Model License. It also lacks the user threshold in Meta’s license. That legal clarity makes Qwen 3.6 attractive for many startups. The main downside is that Qwen comes from Alibaba, a Chinese company. Some US enterprises and agencies have policy restrictions on Chinese-origin AI models. That is a real obstacle, even when the license is permissive. Independent tests show Qwen 3.6 is very strong at multilingual coding and tool selection. It can be weaker than Nemotron 3 Ultra on long-horizon planning and multi-step agent execution.

Hardware requirements are high for the full precision model. But Qwen’s active parameter count keeps latency manageable, and quantized versions are widely used. For teams that want a permissive Apache-licensed coding agent and are not constrained by US procurement rules, Qwen 3.6 remains a top pick. Nemotron 3 Ultra wins on balanced general reasoning and US vendor support.

Key strengths:

  • ✅ Apache 2.0 license is straightforward and commercially friendly
  • ✅ Top-tier coding and agentic tool-use scores
  • ✅ 256K context supports large repository and document tasks
  • ✅ Strong multilingual performance across programming languages and natural languages
  • ❌ Dense variants can require more memory than active-parameter models
  • ❌ Some independent tests show weaker long-horizon planning than Nemotron 3 Ultra
  • ❌ US enterprise buyers may face procurement or policy questions around Chinese origin

Who it’s for: Teams that want a permissive Apache-licensed coding agent and already use Qwen tooling.

Frequently Asked Questions

When did NVIDIA release Nemotron 3 Ultra?

NVIDIA released Nemotron 3 Ultra 550B on June 18, 2026. The weights appeared on Hugging Face the same day. NVIDIA’s official announcement is the primary source. The release includes BF16 and FP8 checkpoints.

What license does Nemotron 3 Ultra use?

The model uses the NVIDIA Open Model License. It permits commercial use, fine-tuning, and distribution of derivative models. It is not fully open-source because the training data and training code are not public. The license also includes defense, surveillance, and export control restrictions.

How many parameters and how much context?

Nemotron 3 Ultra has 550 billion total parameters and 32 billion active parameters per token. The context window is 128K tokens. It supports structured outputs and function calling. Quantized versions reduce memory for shorter contexts.

What hardware do I need to run it?

Full BF16 inference needs roughly 1.1TB of VRAM, typically 8x H100 80GB or 4x H200 141GB. FP8 uses about 550GB of VRAM. AWQ 4-bit can fit on 2x A100 80GB for 32K context. Use vLLM, SGLang, or TensorRT-LLM.

How does it compare to closed models?

NVIDIA reports 85.7 on MMLU Pro, 76.3 on GPQA Diamond, 94.2 on HumanEval, and 93.5 on MATH-500. These scores place it close to Claude Sonnet 4.5 and Gemini 3 Flash on reasoning. It trails specialized closed models on SWE-bench Verified. Independent tests largely confirm NVIDIA’s results.

Is it the best US open-weight model of 2026?

For general reasoning, coding, and commercial local deployment, yes. It beats Llama 4 Maverick on reasoning and matches many closed models. DeepSeek V4 and Qwen 3.6 are stronger in some math and coding benchmarks but are not US-based. The best choice depends on license, hardware, and task.

What Should You Remember?

  • 550B open-weight release: NVIDIA Nemotron 3 Ultra launched June 18, 2026 with 550B total parameters, 32B active, and 128K context.
  • Benchmark parity: NVIDIA reports MMLU Pro 85.7, GPQA 76.3, and HumanEval 94.2, placing the model close to Claude Sonnet 4.5.
  • License: The NVIDIA Open Model License allows commercial use and fine-tuning but is not fully open-source and has defense and surveillance restrictions.
  • Hardware reality: Full BF16 needs roughly 1.1TB of VRAM; FP8 and AWQ checkpoints lower the barrier to multi-GPU systems.
  • US alternative: Nemotron 3 Ultra is the strongest US open-weight generalist, competing with Llama 4 Maverick and non-US DeepSeek V4 and Qwen 3.6.
  • Self-host economics: Teams can avoid per-token API billing by running this model locally on multi-GPU hardware.

Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.