NVIDIA open-sources Nemotron 3 family and training data

Huang Renxun wants a flag in the open model market

While many assumed NVIDIA only sells GPUs, the company quietly released the full Nemotron 3 series — not just model weights, but also the 3 trillion token pre-training dataset, reinforcement learning workflows, and evaluation tools, all open-sourced.

The signal is clear: NVIDIA no longer wants to be just the shovel seller; it wants a place in the open AI ecosystem.

Three sizes, all Mixture-of-Experts

The Nemotron 3 series uses a Hybrid Latent MoE architecture, available in three versions:

ModelTotal parametersActive parameters per inferenceStatus
Nano30 billion~3 billionReleased
Super~100 billion~10 billionFirst half 2026
Ultra~500 billion~50 billionFirst half 2026

The core advantage of MoE is that while total parameters are large, only a fraction is activated per inference, keeping computational costs manageable. DeepSeek V3 and Llama 4 use the same approach, and NVIDIA is now following suit.

Nano benchmark results are interesting

For the released Nano version, official data shows:

  • 4x inference speed over the previous generation
  • 60% reduction in inference token generation (fewer intermediate steps needed to reach conclusions)
  • 1M token context window

Third-party evaluator Artificial Analysis commented: "The most open and efficient model in its size class, with leading accuracy."

This "efficient inference" direction is somewhat similar to DeepSeek V3.2's sparse attention approach — both aim to reduce computation without sacrificing quality.

Fully open toolchain

NVIDIA simultaneously open-sourced a suite of supporting tools:

  • NeMo Gym: Agent reinforcement learning training framework
  • NeMo RL: Reinforcement learning workflow library
  • NeMo Evaluator: Evaluation tool

Model weights, training data, and toolchain are all available on Hugging Face and GitHub under fully open licenses. Anyone can download, fine-tune, and deploy them.

Why NVIDIA is doing this

Several possible rationales:

Software locks in hardware. If you use NVIDIA's open model toolchain, you will likely run it on NVIDIA GPUs. This mirrors the CUDA strategy — using the ecosystem to lock in compute infrastructure.

Hedging against open-source competition. Open models like DeepSeek, Llama 4, and Qwen are getting stronger. If NVIDIA only plays at the hardware layer, its influence in the AI ecosystem will be diluted. Building open models means entering the arena directly.

Enterprise procurement bundling. Many large enterprises buy NVIDIA GPU-plus-service packages. Having trusted open models built in is a selling point and part of the sales material.

Of course, open-sourcing is also a cost for NVIDIA. But given the company's current financial strength, that is not an issue.

Super and Ultra are worth waiting for

The currently released Nano is relatively conservative in scale. What is truly exciting is the Super (100 billion parameters) and Ultra (500 billion parameters) expected in the first half of 2026. If Nemotron 3 Ultra delivers strong benchmarks, this will be more than just "NVIDIA also makes models" — it will be a genuinely competitive player on the open model leaderboard.

Sources: CocoLoop, NVIDIA Debuts Nemotron 3 Family of Open Models (NVIDIA Newsroom)