Nemotron 3.5 Lightning

NVIDIA's 30B-A3B hybrid MoE for one H100: 3B active, 1M context, OpenMDW license. Qwen3.6-35B-A3B scores higher in NVIDIA's own table.

In short. Nemotron 3.5 Lightning is a mid-sized NVIDIA model that answers fast because only 3 billion of its 30 billion parameters work on each word. One 80 GB GPU can run it, it reads long documents, and the permissive license allows commercial use. NVIDIA's own table shows Qwen3.6-35B-A3B ahead of it on most coding tests.

Summary

NVIDIA released Nemotron 3.5 Lightning 30B-A3B on 11 August 2026 as the successor to Nemotron 3 Nano. It is a hybrid mixture-of-experts model that mixes Mamba-2, expert and attention layers, with about 30 billion parameters and 3 billion active per token. It has multi-token-prediction layers, was pretrained on more than 20 trillion tokens with an NVFP4 recipe, and supports up to 1 million tokens of context. A single H100 80 GB serves it with about 256K tokens of context; the full 1M window needs eight H100s or a GB200 or B200. It supports English, Spanish, French, German, Italian and Japanese, and has a switchable reasoning mode.

The BF16 weights are about 66 GB and are meant for post-training, research and building your own quantized versions. For direct deployment NVIDIA points to the NVFP4 build, about 22 GB, which is faster and runs on more hardware. NVIDIA also documents DSpark speculative decoding for low-concurrency data-center serving.

NVIDIA's table compares it with Qwen3.6-35B-A3B, Gemma 4 26B-A4B, Nemotron 3 Nano and Super and GPT-OSS-20B. Lightning scores 51.6 on SWE-bench Verified against 34.1 for Nemotron 3 Nano, but Qwen3.6-35B-A3B scores 70.1, and on Terminal-Bench 2.1 it gets 24.6 against 44.4 for Qwen. It leads only a few rows, such as IFBench among its Nemotron siblings. NVIDIA measured all of this itself, and the license is OpenMDW-1.1, a permissive license that allows commercial use.

Good for

  • Fast agent and chat workloads on a single 80 GB GPU
  • Fine-tuning and distilling a hybrid MoE with open recipes
  • Long-document work, with 256K context on one H100 and up to 1M on larger setups
  • Serving with speculative decoding on Blackwell hardware

Watch out for

  • NVIDIA's own table shows Qwen3.6-35B-A3B well ahead on coding and agentic tests
  • The BF16 repository is meant for customization; use the NVFP4 build to deploy
  • Only six languages are listed as supported
  • Benchmark numbers come from NVIDIA and may differ from other vendors' reports

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Lightning 30B-A3B (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Lightning 30B-A3B (NVFP4) NVFP4 31.6B 3B active 73 GB 36 GB 24 GB Gaming GPU
Lightning 30B-A3B (BF16) BF16 31.6B 3B active 73 GB 36 GB 22 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs.

More about OpenMDW-1.1

See also

Frequently asked questions

What is Nemotron 3.5 Lightning?
Nemotron 3.5 Lightning is a mid-sized NVIDIA model that answers fast because only 3 billion of its 30 billion parameters work on each word. One 80 GB GPU can run it, it reads long documents, and the permissive license allows commercial use. NVIDIA's own table shows Qwen3.6-35B-A3B ahead of it on most coding tests.
Can I use Nemotron 3.5 Lightning commercially?
Yes. Nemotron 3.5 Lightning is released under OpenMDW-1.1, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Nemotron 3.5 Lightning open source?
Nemotron 3.5 Lightning is open-weight, not open source in the strict sense. You can download the model files, but the OpenMDW-1.1 adds conditions that the Open Source Initiative definition does not allow.
Can I run Nemotron 3.5 Lightning on my own computer?
Yes, on a good graphics card. The smallest variant, Lightning 30B-A3B (BF16), needs about 22 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was Nemotron 3.5 Lightning released?
Nemotron 3.5 Lightning was released by NVIDIA on August 11, 2026.
How long a context window does Nemotron 3.5 Lightning have?
Nemotron 3.5 Lightning supports a context window of up to 1M tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Nemotron 3 NVIDIA 31.6B – 120B 12B active 1M NVIDIA Nemotron (Permissive) 21 GB Gaming GPU
Nemotron 3.5 NVIDIA 8B – 120B 12B active 1M OpenMDW-1.1 (Permissive) 5.52 GB Laptop or phone
Qwen3.6 Alibaba Qwen 27.8B – 36B 3B active 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone