Model Drop

Nemotron 3.5 Lightning

August 11, 2026 2 min read

Nemotron 3.5 Lightning, which NVIDIA shipped on August 11, 2026, is an open 30B mixture-of-experts model with 3B active parameters, built for the high-volume execution steps inside long-running agents.

Model: Nemotron 3.5 Lightning (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16)
Maker: NVIDIA
Ship date: August 11, 2026
Pricing: Free (open weights)
Context: Up to 1M tokens
Where you can use it:
- Hugging Face weights in BF16 and NVFP4
- build.nvidia.com and OpenRouter
- Local via LM Studio, llama.cpp and Ollama
When to reach for it:
- Fast tool calls under a bigger planner model
- Fine-tuning on your own domain data
- Skip it for the hardest coding work
How it compares:
- Trails Qwen3.6 35B A3B on most tests
- Beats Gemma 4 26B A4B at web research
- Big step up from Nemotron 3 Nano
What makes it different:
- Open weights, training data and recipes
- Ships with MTP, DSpark and DFlash drafters
Early reactions: Hacker News users praised its speed and fully open recipe, but some noted it trails Qwen3.6 35B A3B and overthinks.
Read more: NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents, NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Deliver Faster, Smarter, More Efficient Agentic AI, Hugging Face model card

The benchmarks

Scores are NVIDIA's own runs of the BF16 weights under one harness (NeMo Gym and NeMo Evaluator) against Qwen3.6 35B A3B, Gemma 4 26B A4B, Nemotron 3 Nano and gpt-oss-20b (NVIDIA also listed the larger Nemotron 3 Super); the card doesn't say which reasoning mode the table used, though reasoning is on by default, and it shows no gpt-oss-20b BrowseComp score.

NVIDIA's numbers

SWE-bench Verified (real GitHub fixes): 51.56%. −18.56 under Qwen3.6 35B A3B (70.12%), −5.84 under Gemma 4 26B A4B (57.40%), +17.48 over Nemotron 3 Nano (34.08%), −0.88 under gpt-oss-20b (52.44%).
Terminal-Bench 2.1 (agentic terminal tasks): 24.58%. −19.80 under Qwen3.6 35B A3B (44.38%), −12.64 under Gemma 4 26B A4B (37.22%), +16.29 over Nemotron 3 Nano (8.29%), +9.41 over gpt-oss-20b (15.17%).
PinchBench (agent task completion): 85.37%. −2.70 under Qwen3.6 35B A3B (88.07%), +10.67 over Gemma 4 26B A4B (74.70%), +19.26 over Nemotron 3 Nano (66.11%), +28.17 over gpt-oss-20b (57.20%).
BrowseComp (hard web research): 36.97%. −11.77 under Qwen3.6 35B A3B (48.74%), +10.67 over Gemma 4 26B A4B (26.30%), +23.23 over Nemotron 3 Nano (13.74%).
GPQA Diamond (PhD-level science, no tools): 75.44%. −7.96 under Qwen3.6 35B A3B (83.40%), −4.17 under Gemma 4 26B A4B (79.61%), +1.39 over Nemotron 3 Nano (74.05%), +3.98 over gpt-oss-20b (71.46%).
IFBench (loose) (instruction following): 71.88%. +8.17 over Qwen3.6 35B A3B (63.71%), −5.37 under Gemma 4 26B A4B (77.25%), −0.29 under Nemotron 3 Nano (72.17%), +3.38 over gpt-oss-20b (68.50%).
AA-LCR (long-context reasoning): 52.00%. −9.06 under Qwen3.6 35B A3B (61.06%), −5.56 under Gemma 4 26B A4B (57.56%), +19.25 over Nemotron 3 Nano (32.75%), +19.12 over gpt-oss-20b (32.88%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 12.9. −5.3 under Qwen3.6 35B A3B (18.2), −3.8 under Gemma 4 26B A4B (16.7), +4.0 over Nemotron 3 Nano (8.9), +3.9 over gpt-oss-20b (high) (9.0).