Overview
NVIDIA-Nemotron-3.5-Lightning-30B-A3B is NVIDIA's throughput-oriented model in the Nemotron 3.5 line, released on August 11, 2026. It is a sparse mixture of experts with 30 billion parameters in total and 3 billion active per token, and the layer stack mixes Mamba-2 blocks with attention rather than using attention everywhere.
| Specification | Nemotron 3.5 Lightning 30B A3B |
|---|---|
| Total parameters | 30B |
| Active parameters | 3B per token, 6 of 128 experts |
| Architecture | Hybrid Mamba-2, MoE and attention |
| Layers | 52 |
| Context window | Up to 1M tokens, 256K on a single H100 |
| Modalities | Text input, text output |
| Reasoning | Switchable through the chat template |
| Speculative decoding | DSpark, MTP and DFlash heads published alongside |
| Release date | August 11, 2026 |
| License | OpenMDW 1.1 |
Mamba-2 layers carry a fixed-size state instead of a KV cache that grows with every token, which is what lets NVIDIA quote a context window up to a million tokens. NVIDIA also publishes DSpark, MTP and DFlash drafter heads for the model, so it can run with speculative decoding without sourcing a separate draft model.
Nemotron 3.5 Lightning benchmarks
NVIDIA's own numbers, from the BF16 model card, compare it with the current same-size open models and with the previous Nemotron generation:
| Nemotron 3.5 Lightning | Qwen3.6-35B-A3B | Gemma 4 26B A4B | Nemotron 3 Super | GPT-OSS 20B | |
|---|---|---|---|---|---|
MMLU Pro Knowledge and reasoning across academic subjects | 81.94 | 85.63 | 85.20 | 83.89 | 76.40 |
GPQA Diamond Expert-level science questions, no tools | 75.44 | 83.40 | 79.61 | 78.60 | 71.46 |
SWE-bench Verified Real-world software engineering tasks | 51.56 | 70.12 | 57.40 | 63.08 | 52.44 |
Terminal-Bench 2.1 Agentic terminal-use tasks | 24.58 | 44.38 | 37.22 | 39.61 | 15.17 |
PinchBench Instruction and formatting robustness | 85.37 | 88.07 | 74.70 | 80.36 | 57.20 |
IFBench (loose) Instruction following | 71.88 | 63.71 | 77.25 | 71.92 | 68.50 |
AA-LCR Long-context reasoning | 52.00 | 61.06 | 57.56 | 58.44 | 32.88 |
On accuracy this is not the leader of its size class: Qwen3.6-35B-A3B is ahead on five of the seven rows, and the coding gap is large. What Lightning is built for is the other axis, serving long contexts fast on one card, and the name says so.
Nemotron 3.5 Lightning hardware requirements
The system requirement to check is memory. We quantized the model from the BF16 reference weights and published the builds as AtomicChat/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF.
| Memory | Build to pick | File size |
|---|---|---|
| 24 GB | AD-IQ4_NL | 19.6 GB |
| 32 GB | Q4_K_M | 25.0 GB |
| 48 GB | Q5_K_M | 26.6 GB |
| 64 GB and up | Q8_0 | 35.0 GB |
The million-token context is a datacenter figure: NVIDIA validates 256K on a single 80 GB H100 and reaches 1M only across eight of them. On a desktop, size the context to the memory left after the weights, and read how the KV cache grows before setting it high.
How to run Nemotron 3.5 Lightning in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for Nemotron-3.5-Lightning in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
The same steps work for every other model in the line, listed on the Nemotron family page. If you would rather drive llama.cpp yourself, our guide to running an LLM locally has the commands.
NVIDIA also ships an NVFP4 build of this model for Blackwell cards; what NVFP4 is explains when that format is worth using over GGUF.
License
The weights ship under the OpenMDW License Agreement 1.1 rather than Apache 2.0 or MIT. NVIDIA states the model is ready for commercial use, and the terms come from the OpenMDW text itself, so read it before you deploy the model in a product.