Nemotron 3.5 Lightning Review

Nemotron 3.5 Lightning is a strong pick for teams that run coding and agent tasks at volume, and a weaker pick for the hardest single-shot reasoning. That is our verdict, and this review shows the reasoning behind it.

The pitch is speed with quality. NVIDIA claims performance near gpt-oss-120b at about a quarter of the total parameters, plus up to 4x faster output than similar-sized models.

We weigh the published benchmarks, the architecture, and the speed claims. We also flag where the model is not the right tool. All benchmark numbers come from NVIDIA's published figures; verify them on the model card.

The Verdict, Up Front

Nemotron 3.5 Lightning earns a recommend for agent and coding workloads that need speed at scale. It is fast, open, and benchmarks well for its active size.

It is not a blanket best-model pick. For the deepest reasoning or nuanced long-form writing, a larger dense model may still win.

If you build agents that run many steps and call many tools, this model belongs on your shortlist. Test it against your own tasks before you commit.

Our score, in short, is high for its lane and clear about its edges. Read on for how each part earns that rating.

What the Benchmarks Show

NVIDIA's published benchmarks put the model in strong company for its active size. These are the company's numbers on the BF16 checkpoint.

On SWE-bench Verified, a real coding-fix test, NVIDIA reports 51.56. That is a serious score for agentic software work.

On GPQA Diamond, a hard science reasoning test, NVIDIA reports 75.44. On MMLU Pro, a broad knowledge test, it reports 81.94. On PinchBench it reports 85.37.

Treat these as NVIDIA's published figures, not independent results. Confirm them on the model card, and test against your own tasks.

Why the Architecture Matters

The architecture is the reason for the speed, and it holds up well. The model interleaves Mamba-2, MoE, and select Attention layers.

Mamba-2 layers handle long sequences cheaply, which suits the 1 million token context. That makes long-context agent work practical, not just possible.

The MoE design gives 30 billion total parameters of knowledge while firing only 3 billion per token. You get breadth without paying dense-model runtime cost.

The select Attention layers keep precise recall where it counts. This mix is why a small active footprint can still benchmark near much larger models.

The Speed Claims, Examined

NVIDIA's speed claims are the headline, and they are specific enough to test. The company reports up to 4x output speed versus similar-sized models.

NVIDIA also claims performance comparable to gpt-oss-120b at about one quarter of the total parameters. If that holds for your tasks, it is a real cost and speed win.

The company adds that the model completes 10,000 tasks about 30 percent faster than Qwen3.6-35B at similar accuracy. That framing targets exactly the agent workloads it is built for.

These are vendor claims measured in vendor conditions. The practical move is to benchmark them on your own workload before you trust them in production.

Strengths That Stand Out

The biggest strength is speed at a strong quality level. A 3B-active model that benchmarks this well changes the cost math for high-volume agents.

Open weights are a second strength. Under the OpenMDW-1.1 license, you can self-host, control your data, and avoid per-token fees.

NVIDIA also released the training data and recipes, which is rare. That openness helps teams that need to audit or adapt the model.

The 1 million token context rounds it out. Long-running agents can hold large state without constant re-fetching.

Where It Falls Short

The model has real weak spots, and speed is part of the reason. A 3B-active footprint favors fast execution over the deepest reasoning.

For the hardest single-shot reasoning or the most nuanced writing, a larger dense model may produce better answers. The benchmarks are strong, but they are not the whole story.

Self-hosting demands GPU capacity and ops skill. That is a barrier for small teams without infrastructure.

Across the model launches we track, the pattern is that speed-tuned task models over-deliver on agent throughput and under-deliver on open-ended judgment. Plan your use around that split.

Coding and Agent Performance

For coding and agent work, this is the model's home turf, and it shows. The SWE-bench Verified score of 51.56 reflects real bug-fixing ability.

SWE-bench Verified tests fixes on real software issues. A score above 50 means the model can carry a fair share of coding tasks end to end.

The 1 million token context helps here too. An agent can load a whole repository and keep the full picture across many steps.

The speed claims matter most in this setting. Faster output means an agent finishes more tasks per hour on the same hardware.

NVIDIA's 30 percent faster claim against Qwen3.6-35B is framed around 10,000 tasks. That volume framing matches how coding agents actually run.

The PinchBench score of 85.37 rounds out the picture. It points to reliable performance on the task suite NVIDIA used.

Reasoning and Knowledge

On reasoning and knowledge, the model is strong but not dominant. NVIDIA reports GPQA Diamond 75.44 and MMLU Pro 81.94.

GPQA Diamond is a hard, graduate-level science test. A score of 75.44 shows solid reasoning for a model this size.

MMLU Pro is a broad, tough knowledge test. An 81.94 signals wide competence across many subjects.

These are good scores for a 3B-active model, driven by the large 20 trillion token training set. Still, a larger dense model may edge ahead on the very hardest single-shot reasoning.

The gap is about the top end, not everyday work. For most reasoning tasks in an agent, these scores are more than enough.

Weigh the scores against speed. A slightly lower reasoning ceiling can be worth it when the model runs several times faster.

Openness and Control

On openness, the model scores near the top of the open field. It ships under the OpenMDW-1.1 license with open weights.

NVIDIA went further than most and released the training data and recipes. That level of openness is uncommon and valuable.

For regulated teams, this control is a real edge. You can self-host, audit behavior, and keep sensitive data on your own hardware.

You also avoid vendor lock-in. If a hosted provider changes terms or price, you can move the same weights elsewhere.

Serving is well supported. Open stacks like vLLM run the model, so you are not tied to one vendor's tooling.

The two checkpoints add flexibility. NVFP4 runs lean on smaller hardware, while BF16 keeps full precision for top accuracy.

Who Should Use It, and Who Should Not

Use Nemotron 3.5 Lightning if you build coding agents or high-volume, multi-step automations. The speed and open license pay off most there.

Use it if you want to self-host and keep data in house. The open weights and released recipes support that fully.

Skip it as your only model if your core need is the deepest reasoning or the most polished long-form writing. Pair it with a larger model for those jobs.

Skip it if you have no infrastructure and only light usage. A small hosted model may cost less for simple, occasional work.

Our Final Word

Nemotron 3.5 Lightning is one of the more compelling open agent models of its generation. The speed-to-quality trade is well judged for real production work.

The published benchmarks are strong for a 3B-active model, and the open license removes lock-in. That combination is hard to match on cost per task.

The real caveat is scope. This is a fast task model, not a universal best-at-everything model, so pair it wisely.

Our recommendation stands: shortlist it for agent and coding workloads, benchmark it on your own tasks, and verify every claim on the official model card.

The launch also fits a broader trend. Open-weight models keep closing the gap with closed ones, and this release pushes that further on speed.

If you already run agents in production, the upside is concrete. More tasks per hour at a similar quality bar can lower your cost per result.

Give it a real trial before you switch anything. A short pilot on your own workload will tell you more than any benchmark table.

For the right team, the payoff is clear and fast. Fast, open, and strong for its size is a rare and useful mix.