NVIDIA's compact execution model for long-running agents, designed to handle high-volume tool calls, result validation, and specialized tasks after larger models perform planning. It has 30B total / 3B active parameters in a hybrid Mamba-2 + Attention MoE: 52 backbone layers interleave Mamba-2, attention, and MoE blocks with 128 routed experts (top-6) plus one shared expert. Multi-Token Prediction (MTP) supplies an additional attention/MoE prediction layer and native speculative decoding.

The model was pretrained from scratch by NVIDIA on more than 20T tokens using an NVFP4 recipe, then continued-pretrained for MTP, supervised fine-tuned, and post-trained with asynchronous multi-environment GRPO across math, code, science, instruction following, multi-step tool use, multi-turn conversation, and structured-output tasks. It supports up to a 1M-token context and can deploy on one 80GB H100 or A100; BF16 reference, NVFP4 inference, pretrained-base, and DSpark/DFlash speculative-drafter checkpoints shipped under OpenMDW-1.1.

NVIDIA's reproducible release evaluation reports MMLU-Pro 81.94, GPQA Diamond 75.44, SWE-Bench Verified 51.56, Terminal-Bench 2.1 24.58, PinchBench 85.37, IFBench 71.88, and AA-LCR 52.0. NVIDIA positions Lightning below larger frontier planners but as a faster, customizable executor for always-on agent workflows, with reasoning enabled or disabled through its chat template.

The technical report covering this model landed with the larger siblings on 2026-10-15, two months after the weights: NVIDIA Nemotron 3.5 Super VL and Lightning Technical Report. The line now spans Super VL at 120B/12B and Bolt at 74B/9B, with Lightning the smallest at 30B/3B.

The joint technical report for Lightning and Super (arXiv 2610.38095, October 28, 2026) describes both as built with large-scale distillation, long-context training and native speculative decoding. Super adds Indic continued pretraining, which NVIDIA says lifts Indian-language benchmarks by up to 9.82 points, and agentic software-engineering mid-training. Its post-training runs SFT, RLVR and RLHF plus specialized teachers and multi-teacher on-policy distillation, which is what turns the text-only base into a multimodal agent. The report gives Super 51.78% on SWE-bench Pro, above Nemotron 3 Ultra with about 4.6× fewer active parameters. MTP boosting after post-training lengthens speculative-decoding acceptance by up to 21.8%, NVFP4 keeps a 99.62% median accuracy recovery against BF16, and serving reaches up to 4,856 output tokens/s per GB200 GPU at 50+ tokens/s per user.

Model Details

Architecture MOE
Parameters 30B
Active params 3B
Experts 128 (top-6)
Context window 1,000,000
Training tokens 20T+
AA Intelligence 6 was 13 on v4.3.2
License OpenMDW-1.1

Benchmark Scores

Benchmark Score Mode
MMLU-Pro 81.94 —
GPQA Diamond 75.44 no tools
SWE-Bench Verified 51.56 —
Terminal-Bench 2.1 24.58 —
PinchBench 85.37 —
IFBench 71.88 loose
AA-LCR 52.0 —

Variants

Name Parameters Notes
Nemotron 3.5 Lightning 30B-A3B BF16 30B Post-trained full-precision reference checkpoint
Nemotron 3.5 Lightning 30B-A3B NVFP4 30B Optimized inference checkpoint
Nemotron 3.5 Lightning 30B-A3B Base BF16 30B Pretrained-only base checkpoint
Nemotron 3.5 Lightning DFlash — External block-diffusion speculative-decoding drafter
Nemotron 3.5 Lightning DSpark — External semi-autoregressive speculative-decoding drafter

Paper

moeopen-weightreasoningagenticefficiencyhybrid-architecture

Related