Nvidia's Nemotron 3.5 Lightning takes aim at the cost floor for always-on agents
The open-weight 30B mixture-of-experts model activates just 3B parameters per token and is built for the repetitive execution work that eats most of an agent's token budget.
Nvidia released Nemotron 3.5 Lightning on Monday, an open 30 billion parameter mixture-of-experts model that activates only 3 billion parameters per token and is aimed at a specific economic problem: the cost of keeping agents running all day.
The pitch, laid out in an Nvidia technical blog, is that modern agent systems are ensembles. A frontier model plans and orchestrates while smaller models grind through tool calls, result validation, code review, and monitoring. That execution layer is where the token volume lives, and Nvidia wants it running on a model whose weights are free. Lightning ships under the permissive OpenMDW-1.1 license with checkpoints on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com, plus an Ollama listing that puts the full 1 million token context version at a 25GB download.
The numbers, all Nvidia's
Nvidia claims up to 4x the output speed of similar-sized open models and roughly 30 percent faster completion of agentic tasks. On PinchBench, the company reports 86 percent accuracy across a 10,000 task run while finishing about 30 percent faster than Qwen3.6 35B at comparable accuracy. A companion post introduces NeMo Switchyard, an open source router that steers each request to the cheapest model that can handle it. Nvidia's internal benchmark for the combination is the headline for pricing watchers: frontier-level accuracy at nearly one third the task completion cost of running Claude Opus 4.8 alone. Every one of those figures comes from Nvidia's own testing, and none has been independently reproduced yet. Treat them as a vendor's opening bid.
What open weights do to per-task math
For GaaS operators, an open execution model changes the shape of the cost curve rather than just its level. API tokens are a variable cost that scales with every loop an agent runs. Self-hosted weights on owned or reserved hardware convert that into amortization, which is exactly the trade our analysis of cheap tokens and costly agents flagged as the industry's next battleground. Lightning runs on a single DGX Spark, on RTX workstations, on Jetson at the edge, and on data center Blackwell and Hopper parts, so the same weights can serve a pilot and a production fleet.
The model slots into agent harnesses including OpenClaw and Hermes Agent through Nvidia's NemoClaw security and management stack, and Nvidia named LangChain, Kong, and Cognition as early integration partners. The obvious comparison is the wave of open Chinese mid-tier models, most recently covered in our look at Qwen's 38 Max preview, which have been the default choice for cost-sensitive execution layers. Nvidia is now competing directly for that slot, with the advantage that it also sells the hardware the weights run on.
The strategic logic is hard to miss. Every agent stack that moves its execution tier to open weights buys fewer API tokens and more GPUs. Whether Lightning's accuracy holds up outside Nvidia's benchmark harness will decide if that trade actually clears, and third-party numbers should arrive within days now that the weights are public.