Nemotron 3.5 Lightning review
A free 30B model from NVIDIA. Only 3B work at a time, so it is fast and reads up to 1M tokens.
A free 30B model from NVIDIA that runs on a single big GPU and reads up to 1 million tokens. It is fast because only 3B parameters work at a time, but on NVIDIA's own table it trails Qwen 3.6 on most tests.
What is Nemotron 3.5 Lightning?
Nemotron 3.5 Lightning is a free AI model from NVIDIA, released on 11 August 2026. It has 30 billion parameters, but only 3 billion work on each word, so it answers quickly. It mixes two designs (Mamba-2 and attention) to handle very long inputs.
Key features
- Free to download, commercial use allowed (NVIDIA OpenMDW license)
- Fast: only 3B of its 30B parameters work per word
- Reads up to 1M tokens (256K on a single H100)
- Thinking mode you can switch on or off
- Ready to run with vLLM, SGLang and Docker, plus smaller community versions for home PCs
Specs and price
| Made by | NVIDIA |
| Released | 11 Aug 2026 |
| Size | 30B parameters, 3B at work per word |
| License | NVIDIA OpenMDW v1.1, commercial use allowed |
| Languages | English and code first, plus Spanish, French, German, Italian and Japanese |
| How much it can read | Up to 1M tokens |
| Computer needed | Full version: 1x H100 or A100 80GB. Smaller NVFP4 and GGUF versions for local devices |
| Download | Hugging Face, nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 |
| Price | Free |
Test scores
| Test | Score | Published by |
|---|---|---|
| MMLU ProGeneral knowledge | 81.94 (Qwen 3.6 35B: 85.63) | NVIDIA |
| GPQA DiamondHard science questions | 75.44 (Qwen 3.6 35B: 83.40) | NVIDIA |
| SWE-bench VerifiedFixing real code bugs | 51.56 (Qwen 3.6 35B: 70.12) | NVIDIA |
| IFBenchFollowing instructions | 71.88 (Qwen 3.6 35B: 63.71) | NVIDIA |
Scores published by NVIDIA are the maker's own numbers. We will add our own results in the hands-on review.
How to download and run Nemotron 3.5 Lightning
- Download it from Hugging Face. NVIDIA recommends the NVFP4 version for everyday use and the BF16 version for fine-tuning.
- On a server: run it with vLLM or SGLang, which give you an OpenAI-style API.
- On your own PC: use a community GGUF version, for example from Unsloth, in LM Studio or llama.cpp.
- Or with Docker:
docker model run hf.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
Who it is for
Good fit
- Developers who want a fast free model for agents
- Teams that want to fine-tune their own model
- Long-document work on one GPU
Skip it if
- Anyone who wants the smartest free model at this size: Qwen 3.6 35B scores higher on most tests
- Non-technical users: there is no simple app
Pros and cons
Pros
- Free, with commercial use allowed
- Very fast for its size
- Reads up to 1M tokens
- Many ways to run it
Cons
- Trails Qwen 3.6 35B on knowledge, science and coding
- Full 1M context needs 8 GPUs
- NVIDIA says the main version is meant for customising, not direct use
- No simple consumer app
How it compares
On NVIDIA's own table, Qwen 3.6 35B beats it on general knowledge, science and coding. Lightning wins on following instructions (71.88 against 63.71). Its real advantage is speed and long memory on one GPU.
Hands-on review In testing
We are running Nemotron 3.5 Lightning on real tasks. Results will be added to this page, at the same address.
How it got here
| Date | What happened | Source type |
|---|---|---|
| Aug 11, 2026 | Nemotron 3.5 Lightning released on Hugging Face.NVIDIA model card on Hugging Face | Confirmed |
Frequently asked questions
Is Nemotron 3.5 Lightning free?
Yes. You can download it from Hugging Face for free, and NVIDIA's license allows commercial use.
Can I run Nemotron 3.5 Lightning on my PC?
The full version needs an 80 GB data-centre GPU. Smaller community versions (GGUF) and NVIDIA's NVFP4 version are made for local devices.
How much text can it read?
Up to 1 million tokens. On a single H100 it is set up for 256,000.
Is Nemotron 3.5 Lightning better than Qwen 3.6?
Not on most tests in NVIDIA's own table. It wins on following instructions and is very fast.
Related models
Flint
RecursiveFlint is out. Launch facts from official sources are below, and our hands-on review follows.
Kolibri
Aleph AlphaRuns cheaply on-premise, which companies with strict data rules need.
Muse Glimmer
MetaMeta's only free model this year. Many people searching for Llama 5 actually want this.