On August 11, NVIDIA put Nemotron 3.5 Lightning (30B-A3B) on Hugging Face and Build.NVIDIA.com under the OpenMDW v1.1 license, cleared for commercial use. The Nemotron family has always had a clear positioning — open weights, open training data, open recipes — and the Lightning tier targets the "sub-agent workhorse": long-running autonomous agents, sub-agent deployments, and agentic workflows. The Build.NVIDIA.com page tags it as "the fastest 30B A3B MoE model" — that's NVIDIA's own claim, but combined with the 3B-active-parameter design intent, the direction is clear: models in this class don't chase single-step reasoning leaderboard wins; they aim to absorb high-frequency calls cheaply and reliably.
Architecture: Three Components Interleaved
Lightning 3.5 uses a hybrid Mixture-of-Experts architecture: Mamba-2 layers, MoE layers, and select Attention layers are interleaved, with 30B total parameters and 3B active per token, plus a context window of up to 1M tokens. It supports English and coding languages, along with Spanish, French, German, Italian, and Japanese. Recommended sampling parameters are stated right in the model card: Temperature 1.0, Top_P 0.95.
Training: Five Stages, from NVFP4 Pre-training to PTQ
The model card's training methodology section breaks down into five stages, each with verifiable details:
Stage 1, pre-training: more than 20 trillion tokens, trained with an NVFP4 recipe — note this is "pre-training in a low-precision format," not post-training quantization. Pre-training data cutoff is September 2025; post-training data cutoff is May 2026. The software stack is Megatron-LM.
Stage 2, continued pre-training for MTP: the model gains Multi-Token Prediction layers that predict multiple future tokens simultaneously, providing richer training signals.
Stage 3, SFT: synthetic code, math, science, tool-calling, instruction-following, and structured-output data, plus tasks designed for long-range retrieval and multi-document aggregation.
Stage 4, reinforcement learning: multi-environment RL using GRPO (Group Relative Policy Optimization) across math, code, science, instruction following, multi-step tool use, multi-turn conversation, and structured-output environments, with an asynchronous architecture that decouples training from inference and leverages MTP to accelerate rollout generation.
Stage 5, PTQ: post-training quantization with NVIDIA Model Optimizer — a "Four Over Six NVFP4" variant of static MSE calibration, W4A16 on routed and shared experts, FP8 per-tensor dynamic scales on Mamba projections and KV cache, calibrated with 1000 samples at 32k token length.
One thing worth highlighting is the transparency of post-training synthetic data: the model card lists the teacher models — including DeepSeek-V4-Pro, Nemotron 5.5, GPT-5.5, GLM-5, and gpt-oss-120b — so who distilled from whom is fully visible.
Benchmarks: BF16 and NVFP4 Nearly Neck and Neck
The official evaluation table gives both BF16 and NVFP4 columns, which is far more valuable than a single column of numbers:
| Benchmark | BF16 | NVFP4 |
|---|---|---|
| MMLU Pro | 81.94 | 81.62 |
| GPQA Diamond (no tools) | 75.44 | 75.57 |
| SWE-bench Verified | 51.56 | 52.80 |
| Terminal-Bench 2.1 | 24.58 | 23.46 |
| BrowseComp | 36.97 | 36.81 |
| AA-LCR (long context) | 52.00 | 49.19 |
On most benchmarks NVFP4 lands within 1 point of BF16, and SWE-bench Verified even comes out ahead (52.80 vs 51.56); long-context AA-LCR drops a bit more (52.00 to 49.19). Evaluation recipes, containers, prompts, and scoring configurations are all published in the NeMo Gym repository for reproducibility. SWE-bench and Terminal-Bench used NeMo Evaluator.
On the Deployment Side
Hardware compatibility spans three tiers: Blackwell, Hopper (NVFP4/W4A16), and Ampere (W4A16). Inference engine: Dynamo + vLLM, tested on H100. The NGC catalog also hosts a BF16 full-precision version, which the official notes say is intended primarily for customization and post-training rather than direct production inference — the production path defaults to the quantized build.
So What
The most memorable things about Nemotron 3.5 Lightning aren't the benchmark numbers, but two structural facts. First, NVFP4 goes directly into the pre-training recipe — quantization and training are no longer two disconnected processes. Second, the model card lays everything bare: the 20T pre-training corpus, synthetic-data teacher models, RL environments, quantization calibration parameters. On the "openness" front, NVIDIA has raised the bar to "open enough to reproduce." Developers building agentic systems can treat it as a public engineering reference ready to use as a sub-agent foundation.
(Source: https://build.nvidia.com/nvidia/nemotron-3.5-lightning-30b-a3b/modelcard )