NVIDIA's Nemotron 3.5 Lightning: a 30B open MoE for always-on agents
3B active params, a 1M-token window and ~670 tok/s under a commercial-open license — self-host it in Ollama, or route agent grunt-work to it via Switchyard.
Copy markdown
3B active, but it plays in the 30B league
Nemotron 3.5 Lightning is an open Mixture-of-Experts model — 30B total, only 3B active per token — with a 1M-token context and ~670 tokens/sec, up to 4x the output speed of similar-sized models. NVIDIA measures 86% on PinchBench while finishing 10,000 tasks 30% faster than Qwen3.6 35B.
OpenMDW license: ship it commercially, no asterisks
The weights land under the permissive OpenMDW-1.1 license — open for commercial use without material restrictions — so there's no research-only clause to trip over. NVFP4 and BF16 checkpoints are already on Hugging Face, ModelScope, and build.nvidia.com.
Run it tonight in Ollama, or skip hosting entirely
GGUF conversions work in LM Studio, llama.cpp, Ollama and Unsloth, making a fast local agent loop a same-evening build. Don't want to host? Serverless inference is already live on DeepInfra, Fireworks, FriendliAI, CoreWeave, Nebius and Crusoe.
NeMo Switchyard: cheap model for grunt work, frontier for planning
NVIDIA shipped a routing library alongside the model: Switchyard sends high-volume execution to Lightning and routes hard planning to a frontier model, trimming token spend on long-running agents. It's the 'small model does the typing' pattern, packaged.
Elsewhere: Copilot for JetBrains now speaks Ollama
GitHub Copilot added Ollama as a bring-your-own-key provider inside JetBrains IDEs, plus cross-session memory — so you can point Copilot at a local model like Lightning and stop re-explaining your project every chat.