NVIDIA Open-Sources Nemotron 3 Ultra, a 550B Mamba-Transformer Mixture-of-Experts Built for Long-Running Agents
The 550-billion-parameter model activates 55 billion parameters per token, ships under the Linux Foundation's OpenMDW-1.1 license, and trades blows with China's Kimi-K2.6 on coding benchmarks.
Overview
NVIDIA has released Nemotron 3 Ultra, a 550-billion-parameter open-weights language model that activates only a fraction of its parameters on each token and is aimed squarely at long-running autonomous agents. According to MarkTechPost, the model is a 550 billion total parameter Mixture-of-Experts (MoE) system in which only 55 billion parameters are active per token. The model became available on June 4, 2026, with AWS announcing day-zero availability on Amazon SageMaker JumpStart the same day.
The release is the largest entry to date in NVIDIA’s Nemotron line of open models, following the 30-billion-parameter Nemotron-Cascade 2 reasoning model the company open-sourced in March.
What We Know
Architecture. Nemotron 3 Ultra departs from a conventional Transformer. AWS describes it as a hybrid Transformer-Mamba Mixture-of-Experts (MoE) architecture that activates only 55B of its 550B parameters per forward pass. MarkTechPost adds that the model uses a hybrid Mamba-Attention architecture instead of a pure Transformer, with 108 layers and a model dimension of 8,192, and that it routes tokens across 512 experts, with the top 22 activated per token.
Precision and training. The model was trained in NVIDIA’s 4-bit NVFP4 format. According to MarkTechPost, the NVFP4 pre-training uses an E2M1 4-bit datatype with two-dimensional block quantization.
Context window. The model handles up to 1M tokens, according to AWS. MarkTechPost reports that context was extended to 1 million tokens and that the model scores 94.7 on the RULER long-context benchmark at 1 million tokens. The model is text in, text out, per AWS.
Efficiency claims. AWS cites 5x faster inference for agent workflows and up to 30% lower cost for agentic workloads. MarkTechPost puts more precise numbers on the throughput claim: on an 8K input / 64K output setting at NVFP4 on GB200, Nemotron 3 Ultra reaches 5.9x the throughput of GLM-5.1 and is 4.8x faster than Kimi-K2.6.
Benchmark standing. The comparison with China’s open-weights frontier is not one-directional. According to MarkTechPost, on Terminal Bench Kimi-K2.6 leads at 67.2 while Nemotron 3 Ultra scores 56.4 — so even as NVIDIA’s model runs faster, the Chinese model holds the higher accuracy on that agentic-tool benchmark.
Licensing. Both AWS and MarkTechPost confirm the model is fully open. MarkTechPost identifies the license as OpenMDW-1.1, the Linux Foundation’s permissive license for open model distributions.
Availability. Beyond Amazon SageMaker JumpStart, MarkTechPost reports the model is available via Hugging Face, NVIDIA NIM, Nebius Token Factory, OpenRouter, Together AI, and Perplexity.
What We Don’t Know
NVIDIA’s published throughput and cost figures are vendor benchmarks measured on its own GB200 hardware, and independent third-party reproductions of the 5.9x and 4.8x throughput multipliers were not available at release. The relationship between the model’s accuracy and NVIDIA’s NVFP4 precision format — and how performance holds up on non-NVIDIA accelerators — has not been independently characterized in the sources reviewed. Pricing for the hosted endpoints across the various providers was also not detailed at launch.
Analysis
Nemotron 3 Ultra extends a pattern visible across NVIDIA’s recent open-model releases: rather than chasing the single highest benchmark score, the company is optimizing the cost and speed of running a model long enough to complete multi-step agentic tasks. The architectural choices reflect that goal. A Mamba-Transformer hybrid trades some of the pure-attention Transformer’s quadratic cost for the linear-time behavior of state-space layers on long sequences, and a Mixture-of-Experts design keeps the active parameter count — and therefore the per-token compute — far below the 550-billion total, with only 55 billion live on any given token according to AWS.
That efficiency framing is also where the competitive story sits. On the throughput numbers NVIDIA reports, Nemotron 3 Ultra is faster than both a Western open model (GLM-5.1) and the Chinese Kimi-K2.6, per MarkTechPost — yet on the Terminal Bench accuracy figure in the same report, Kimi-K2.6 still leads. For teams weighing open-weights options for agent workloads, that is the tradeoff to watch: speed and an open license against raw benchmark accuracy.