NVIDIA's Nemotron 3.5 Lightning: open agent LLM you run locally
A 31.6B/3.6B-active MoE that ties gpt-oss-120b, ships under a commercial-OK open license, runs in Ollama, and bills $0.05/$0.20 per M serverless.
Copy markdown
gpt-oss-120b brains, a quarter the size
Nemotron 3.5 Lightning is a 31.6B-total / 3.6B-active MoE that ties gpt-oss-120b on Artificial Analysis's Intelligence Index (24) and beats it on GDPval-AA v2 (824). Frontier-ish reasoning without frontier-sized hardware.
It fits on the GPU you already own
Weights are out under the permissive OpenMDW-1.1 license — commercial use is fine — with a Q4_K_M GGUF near 17–20GB that runs in llama.cpp, Ollama, and LM Studio on a single 24GB card. Full BF16 wants one H100/A100 80GB.
Built to grind, not to chat
It's tuned for the boring agent loop — tool calls, result validation, subagent delegation — hitting ~670 tokens/sec and finishing agentic workloads up to 30% faster than similar models at ~86% PinchBench accuracy. Text-only, 1M-token context.
Serverless is basically free
Don't want to self-host? It's live day-zero on DeepInfra at $0.05 in / $0.20 out per million tokens, plus Fireworks, CoreWeave, Nebius, Crusoe and build.nvidia.com. At that price, agent runs that used to cost real money become rounding errors.