NVIDIA's Nemotron 3.5 Lightning: 30B open agent model + router
Runs on your desk or $0.08/M on OpenRouter; the open Switchyard router cut one partner's agent bill 58% while holding frontier accuracy.
Copy markdown
A 30B open agent model that runs on your desk
Nemotron 3.5 Lightning is a 30B-total / 3B-active MoE (Mamba-2 + attention hybrid) under the commercial-friendly OpenMDW-1.1 license. It runs locally on RTX PCs, DGX Spark, or Jetson, or on vLLM/SGLang, and GGUF quants are already up for llama.cpp.
4x faster tokens, ~30% faster agent runs
NVIDIA claims up to 4x faster output and ~30% faster agentic task completion versus models in its class, at 51.6% on SWE-Bench Verified and 81.9% on MMLU Pro. For long-running agent loops, that is real wall-clock and token savings.
Switchyard routes your agent to the cheapest model that works
NeMo Switchyard is an open-source routing library that sends each step to the best model by quality, latency, or cost. NVIDIA's internal tests hold frontier accuracy at roughly one-third the cost of Opus 4.8 alone. Build a multi-model agent this weekend that stops paying flagship prices for easy steps.
The savings are already landing in production
Early users report concrete cuts from routing instead of defaulting to one model: Ramp shaved 58% off cost and 33% off runtime, Cognition 28%, and Classmethod 27%.
Cheap on OpenRouter, or free to self-host
Don't want to run it yourself? It's live on OpenRouter at $0.08/M input and $0.20/M output with a 1M-token context, a fraction of frontier pricing. The weights are open if you'd rather keep everything in-house.
Elsewhere: GitHub Copilot runs local models via Ollama in JetBrains
Copilot for JetBrains now supports Ollama plus persistent memory, so you can point it at a local open model like Nemotron and keep your code on your own machine.