Nemotron 3.5 Lightning
NVIDIA's 30B-A3B open hybrid MoE for always-on agents — 1M context, open weights, data, and recipes.
About
Nemotron 3.5 Lightning (August 11, 2026) is NVIDIA's smallest Nemotron 3 family member: 30B total parameters with only 3B active per token, built on a hybrid of interleaved Mamba-2 and MoE layers plus select attention layers, trained on 20T+ tokens. NVIDIA positions it as the execution layer of 'always-on' agents — the model handling routine high-volume work like tool calls, output validation, and delegation, explicitly including code review bots. The release stands out for openness: weights, training data, and recipes under the permissive OpenMDW-1.1 license, which allows commercial retraining and redistribution without crediting or paying NVIDIA. Ships in BF16, FP8, and NVFP4 with speculative-decoding support. Caveat: NVIDIA's 'up to 4x faster output than comparably sized systems' claim had no independent verification published as of late August 2026.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Up to 4x faster output than comparably sized systems — provider-reported, no independent verification published as of late Aug 2026
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Aug 11, 2026 release date consistent across Tech-Insider, Akash Network guide, and HF model-card mirrors; 30B-A3B hybrid Mamba-2/MoE, 1M context, OpenMDW-1.1 license (weights+data+recipes) well documented; '4x faster' claim provider-reported with no independent verification as of late Aug 2026 — labeled as such.