NVIDIA Nemotron 3.5 Lightning (30B-A3B)

2026-08-11 · Publisher : NVIDIA · Domain : Artificial Intelligence

NVIDIA releases Nemotron 3.5 Lightning, an open MoE with 30 billion parameters, 3 billion of them active, designed for fast agent execution and for specialization through post-training, under the OpenMDW-1.1 license.

NVIDIA releases Nemotron 3.5 Lightning, an open MoE with 30 billion parameters, 3 billion of them active, designed for fast agent execution and for specialization through post-training, under the OpenMDW-1.1 license. ### Context Nemotron 3 Nano dates from December 2025. NVIDIA is offering a 3.5 version here that emphasizes execution speed and the ability for a company to regenerate its data and specialize the model at low cost. The launch comes with NeMo Switchyard. ### What the model brings - A claimed throughput of up to 4 times that of comparably sized models, and 10,000 PinchBench tasks run 30% faster than Qwen3.6 35B at comparable accuracy. - Multi-token prediction built in from pretraining, with DFlash and DSpark speculative decoding drafts provided. - Pretraining on more than 20 trillion tokens, in four stages: NVFP4 pretraining, continuation for multi-token prediction, supervised fine-tuning on synthetic data, and multi-environment GRPO reinforcement. - Open datasets, including the Nemotron-CC-v2.1 collection (9.1 trillion web tokens) and the Nemotron-RL Agentic Terminal Pivot dataset. - A context of up to 1 million tokens and the ability to run on a single GPU. ### Architecture and specifications 30 billion parameters in total, 3 billion active, Mamba-2, MoE and attention, context of up to 1 million tokens. License: OpenMDW-1.1 (weights, training data and recipes published "as permissively as possible" according to NVIDIA). Formats: BF16, NVFP4, BF16 base. Openness: weights, pretraining and post-training data, RL environments and recipes (NeMo Automodel, NeMo RL, NeMo Gym); the full corpus is not guaranteed to be published. Compatible with vLLM, SGLang, TensorRT-LLM, Ollama and llama.cpp. ### Cost and access | Item | Detail | |---|---| | Official API | Not disclosed (several hosts) | | Hardware | 1 x H100 for 256K of context, up to 8 x H100 for 1M; GB200 and B200 with speculative decoding; DGX Spark, Jetson, RTX 5090 (NVIDIA post) | | Weights | BF16, NVFP4, base and DFlash/DSpark drafts on Hugging Face (August 1 to 5, 2026) | ### Benchmarks | Benchmark (HF card, NVIDIA) | Lightning | Qwen 3.6 35B A3B | Gemma 4 26B A4B | Nemotron 3 Nano | Nemotron 3 Super | |---|---|---|---|---|---| | MMLU-Pro | 81.94 | 85.63 | 85.20 | 78.46 | 83.89 | | GPQA Diamond | 75.44 | 83.40 | 79.61 | 74.05 | 78.60 | | SWE-bench Verified | 51.56 | 70.12 | 57.40 | 34.08 | 63.08 | | Terminal-Bench 2.1 | 24.58 | 44.38 | 37.22 | 8.29 | 39.61 | | IFBench (loose) | 71.88 | 63.71 | 77.25 | 72.17 | 71.92 | | AA-LCR | 52.00 | 61.06 | 57.56 | 32.75 | 58.44 | Source: Hugging Face card nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16, NVIDIA measurements; all comparators are open. ### Versus closed models No closed model appears in NVIDIA's table. On these tests, Lightning trails Qwen 3.6 35B A3B, notably on SWE-bench Verified (51.56 vs 70.12) and Terminal-Bench 2.1 (24.58 vs 44.38), and is ahead of it only on IFBench (71.88 vs 63.71). Its claimed value is speed (up to 4 times) and the cost of specialization, not the highest level of accuracy. NVIDIA states that the model defines the accuracy-speed frontier on the Artificial Analysis Intelligence Index, a claim not verified at the source. No comparison with a contemporary closed model is established here. ### Key takeaways Lightning targets fast agents and reproducible specialization, with a more permissive license than that of the Nano generation. Its scores on agentic coding tasks remain below those of an open competitor of the same size according to NVIDIA.

Tags : LLM, Open Source, MoE, NVIDIA, Nemotron 3.5, OpenMDW-1.1, Poids, données et recettes ouverts