NVIDIA's Nemotron 3.5 Lightning: open 30B agent model on one GPU
Commercial OpenMDW license, 3B active params, a 1M-token window and ~670 tok/s — plus NeMo Switchyard routing that cut agent cost 74% in LangChain tests.
Copy markdown
One H100, 3B active, 1M context
It's a 30B mixture-of-experts but fires only 3B params per token, so it runs on a single H100 or DGX Spark. The 1M-token window and ~670 tok/s make it up to 4x faster than similar-size open models.
Open weights, data and recipes — commercial-ready
Ships under the permissive OpenMDW-1.1 license with the weights, training recipes, and the released Nemotron-RL-Agentic-Terminal-Pivot dataset. You can fine-tune and ship it in products, not just demo it.
NeMo Switchyard cut agent cost 74%
A companion open-source library routes each step of an agent workflow to the cheapest capable model. In LangChain testing it sent just 7% of calls to a frontier model and cut cost 74% versus a frontier-only baseline.
Tuned for terminal and coding agents
SWE-bench Verified around 52%, Terminal-Bench 24% (over 3x its Nano predecessor), and a GDPval agentic Elo of 824 that beats gpt-oss-120b. It's built for always-on, high-volume agent loops.
The catch: raw reasoning trails the pack
On the Artificial Analysis Intelligence Index it scores 24 — level with gpt-oss-120b but behind Qwen3.6 35B (32) and Muse Glimmer (35). Reach for it for speed and cost, not peak reasoning.