NVIDIA released Nemotron 3.5 Lightning, an open-weight 30-billion-parameter model in a compressed 4-bit form that runs on a single H100 or a DGX Spark desktop machine. It handles up to one million tokens of text at once and is aimed at long-running agents, chatbots and coding tasks, with published benchmark scores and reproducible test recipes. Use requires NVIDIA GPUs and Linux with one of three supported serving frameworks.
What changed
Earlier Nemotron Lightning releases did not ship a compressed 4-bit checkpoint tuned for single-GPU and desktop-class NVIDIA hardware with this context length.
What it unlocks
Running a million-token-context reasoning and tool-calling model on one H100 or a single DGX Spark desktop machine, with open weights and commercial use permitted.
- 30B total parameters, 3B active
- up to 1M token context length
- runs on 1× DGX Spark (GB10) or 1× H100
- SWE-bench Verified 52.80
- pre-trained on over 20T tokens
What you need to act on it
- an NVIDIA Blackwell, Hopper or Ampere GPU
- Linux and a supported serving stack (vLLM, TensorRT-LLM or SGLang)
- acceptance of the OpenMDW-1.1 license
- huggingface.co2026-08-11