Nvidia Releases Nemotron 3.5 Lightning: A 31.6B-Parameter Open Model Built for Speed, Not Peak Intelligence

Nvidia has released Nemotron 3.5 Lightning, the first model in its new Nemotron 3.5 lineup, and the numbers show a company betting on inference speed rather than chasing the top of intelligence leaderboards.

The model directly succeeds the Nemotron 3 Nano 30B A3B and retains its hybrid Mamba-Transformer architecture. It has 31.6 billion total parameters but activates only 3.6 billion at any given time, keeping compute costs low relative to its size class.

Benchmark Performance

According to independent benchmarking platform Artificial Analysis, Nemotron 3.5 Lightning scores 24 on the Intelligence Index — a nine-point jump from its predecessor's score of 15. That ties it with OpenAI's gpt-oss-120b (also 24), despite Lightning running with roughly a quarter of the active parameters. It sits just behind Nvidia's own Nemotron 3 Super (26), a model about four times larger.

Other small models in the same class still lead on raw intelligence: Qwen3.6 35B A3B scores 32, and Meta's Muse Glimmer reaches 35 on the same index.

Where Lightning stands out is speed. In pre-release tests using the final NVFP4 weights, Nvidia claims the model hits nearly 670 tokens per second — the highest measured throughput among all models Artificial Analysis compared, and almost twice the rate of Google's Gemini 3.5 Flash-Lite (386 tokens/s). A typical Intelligence Index task takes about 0.5 minutes for Lightning to complete, versus roughly 3.5 minutes for Qwen3.6 35B A3B and 5.8 minutes for Gemma 4 31B.

Proprietary models still dominate the top of the efficiency frontier overall. Gemini 3.5 Flash-Lite scores 37 on the Intelligence Index at a similar time-per-task, and GPT-5.6 Luna (max) reaches 52 points in under two minutes.

Agentic Gains

The largest improvements appear on agentic benchmarks. On GDPval-AA v2, Lightning posts an Elo rating of 824 — a 334-point gain over Nemotron 3 Nano — surpassing both gpt-oss-120b (800) and the larger Nemotron 3 Super (698). On Terminal-Bench v2.1, its score jumps from 7 percent to 24.3 percent, nearly matching gpt-oss-120b's 26.2 percent.

Artificial Analysis reports that Nvidia worked with partners including CodeRabbit and Harvey during post-training to strengthen performance in specific domains, likely contributing to the agentic-task gains.

Availability and Licensing

Nvidia ships Nemotron 3.5 Lightning under the permissive OpenMDW-1.1 license, positioning it as a high-throughput option for agent-based pipelines rather than a general chatbot replacement. Weights are available now in both BF16 and NVFP4 formats; according to Artificial Analysis, the NVFP4 variant also scores 24 on the Intelligence Index with minimal quality loss versus the higher-precision version.

The model is text-only and supports a one-million-token context window. Serverless inference is already available through DeepInfra, Fireworks, FriendliAI, CoreWeave, GMI Cloud, Nebius, and Crusoe, among others. Pricing for these hosted endpoints was not disclosed in Nvidia's release materials.

What This Means

Nvidia's efficiency argument isn't new — a widely discussed company paper last year claimed models under 10 billion parameters can handle most agent workloads as well as 70- to 175-billion-parameter models, at one-tenth to one-thirtieth the cost. Nemotron 3.5 Lightning, with 3.6 billion active parameters beating a 120-billion-parameter model on agentic benchmarks, is the clearest product-level test of that thesis so far. For teams building agent pipelines where throughput and cost per task matter more than peak reasoning ability, Lightning offers a genuine alternative to larger open models like gpt-oss-120b. It won't compete with frontier proprietary systems like GPT-5.6 Luna on raw intelligence, but that was never the goal — Nvidia is optimizing for a different point on the cost-speed-intelligence curve, and the agentic benchmark numbers suggest that bet is paying off.