Nvidia's Nemotron 3.5 Lightning: a 30B open MoE that runs on one H100
The weights, training data and recipes all ship open; it clears ~52% on SWE-bench Verified, and a companion NeMo router cuts agent cost to a third.
Copy markdown
30B total, 3B active — and it runs on one H100
A hybrid Mamba-2 + MoE model with a 1M-token context and up to 4x faster output than same-size rivals. It scores ~82 on MMLU Pro and ~52% on SWE-bench Verified — roughly gpt-oss-120b quality at a quarter of the parameters. Ollama, LM Studio and llama.cpp all support it on day one.
The license is the real unlock
Weights, training data and recipes all ship under the permissive OpenMDW-1.1 license, cleared for commercial use. Build this weekend: a fully local 1M-context coding agent on a single RTX box, or fine-tune it on your own data with NeMo — no per-token bill.
NeMo Switchyard routes agents to the cheapest model that works
An open-source router shipped alongside it sends each step of an agent workflow to the smallest capable model, holding frontier-level accuracy at about a third the cost of running Opus 4.8 alone — 74% cheaper in Nvidia's own tests. It's on GitHub now.
Elsewhere: GitHub Copilot for JetBrains adds local models
Copilot now accepts Ollama as a BYOK provider in JetBrains IDEs, so you can point it at local models, plus a new cross-session memory that recalls project context between chats. Both switch on in the latest plugin.