Nvidia's new open model has 30 billion parameters and uses 3 of them at a time — and it runs on a desktop
Nvidia released Nemotron 3.5 Lightning on 11 August: a 30-billion-parameter mixture-of-experts model that activates only 3 billion parameters per token, with a context window of up to 1 million tokens and weights under the permissive OpenMDW-1.1 licence — companies can download, modify and build on it without asking Nvidia. It is built for the unglamorous half of agent work: code review, tool calls, watching security alerts, answering billing questions, the steps a frontier model would plan but need not perform itself. Nvidia claims up to four times the output speed and 30% faster agentic task completion than models of similar size, and reports 81.62 on MMLU Pro, 75.57 on GPQA Diamond and 52.80 on SWE-bench Verified. It runs locally on RTX PCs, DGX Spark, DGX Station and Jetson boards as well as in data centres, and is available on Hugging Face, ModelScope and OpenRouter. Alongside it Nvidia open-sourced NeMo Switchyard, a routing library that sends each step of a workflow to the cheapest model that can handle it; Nvidia's own benchmarks put a Switchyard-routed setup at close to a third the cost of running Opus 4.8 for everything.
Related stories
Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship2026-08-21DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21ChatGPT can now read your iMessages and send them — and the setting that lets it skip asking is the one OpenAI warns about
2026-08-21