A 30B mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → with 3B active parameters and 1M context is now free in Cline, aimed at the many fast read-edit-test calls in an agentic loopThe cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.Full definition →. Throughput claims (up to 4x) are NVIDIA's preliminary numbers.
Key takeaways · AI-distilled
The post reports Nemotron 3.5 Lightning scoring 86.2 on PinchBench, 54.3 on SWE-benchThe standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.Full definition →, and 73 on AA-Omniscience non-hallucinationWhen a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.Full definition → despite only 3B active parameters.
Lightning is distilled from NVIDIA's larger Nemotron 3 Ultra and is described as open and customizable: it can be post-trained for specialized workflows and run locally, at the edge, in a datacenter, or in the cloud.
To use it in Cline, pick Cline Usage-Billing as the API provider in Settings, then choose nvidia/nemotron-3.5-lightning from the free model dropdown; a free Cline account and the IDE extension or CLI are required.
Terms in this piece · Glossary
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
agentic loop — The cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
