Nemotron 3.5 Lightning
NVIDIA's 293-tok/s execution-tier MoE ships weights, data, AND recipes — under OpenMDW again
What it is
A 30B-total / 3B-active hybrid that interleaves Mamba-2 and mixture-of-experts layers with select attention: 52 layers, 6 of 128 routed experts plus one shared expert per token, 1M-token context. Five checkpoints shipped at once — BF16 reference, NVFP4 deployment, the pre-trained Base, and two speculative-decoding draft variants — alongside the actual pre- and post-training datasets (Nemotron-CC-v2 at 9.1T tokens among them) and training recipes.
Why it's interesting
NVIDIA is refreshingly honest about what this is: an execution-tier model for tool calls and validation loops, not a peak-accuracy play — its own table shows Qwen 3.6 35B ahead on 13 of 14 rows while Lightning does 293 output tok/s, third-fastest of 134 models tracked by Artificial Analysis. And the release radius is the real story: weights, data, and recipes together is a level of reproducibility the big permissive-license labs still don't offer.
Use cases
- High-throughput agent execution loops — tool calls, validation, routing
- 1M-token context processing where speed matters more than peak accuracy
- Reproducibility research: retrain or ablate against released data and recipes
Who it's for
Agent-infra teams optimizing cost per call; open-training researchers
Setup
Moderate. ~60GB BF16 or the NVFP4 build for single-GPU serving; day-one GGUF via ggml-org, unsloth, and bartowski for local use
Limitations & cautions
It loses to same-class Qwen on nearly every quality benchmark by NVIDIA's own table — this is a speed play, full stop; most of the released datasets sit behind HF gating and approval; and OpenMDW, however permissive, still isn't the three-letter license your counsel already knows.
Editorial takeaway
The most reproducible release of the week carries the asterisk, while less-forthcoming labs ship clean MIT. Our bright line stays put — but this is the entry that makes it feel expensive.