AI
Nvidia releases 30B Nemotron model with 1M context
Nvidia released Nemotron 3.5 Lightning, a 30B open model with 3B active parameters, 1M-token context and NVFP4 deployment support.
Nvidia has released Nemotron 3.5 Lightning 30B-A3B-NVFP4, an open-weight language model with 30 billion total parameters but only 3 billion active per token. According to Hugging Face, the model is available through Hugging Face under Nvidia’s OpenMDW-1.1 license and is marked ready for commercial use.
The model combines Mamba-2, Mixture-of-Experts and Transformer attention layers, and supports context windows of up to 1 million tokens. Nvidia says it was pretrained on more than 20 trillion tokens using crawled and synthetic code, math, science and general-knowledge data, with a pretraining cutoff of September 2025 and post-training data cutoff of May 2026.
It is designed for long-running autonomous agents, sub-agent deployments and local inference. Supported languages are English, Spanish, French, German, Italian and Japanese, alongside coding languages. The model accepts text input and produces text output, and is intended for agent systems, chatbots, retrieval-augmented generation and other AI applications.
Deployment targets and decoding methods
Nvidia lists single-GPU deployment on one DGX Spark with the GB10 chip or one H100. Supported hardware also includes Blackwell systems such as GB200 and GeForce RTX 5090, Hopper GPUs including H100 and H200, and Ampere GPUs through W4A16 kernels.
The NVFP4 checkpoint uses Nvidia’s post-training quantization recipe, with W4A16 for routed and shared experts, FP8 dynamic scaling for Mamba projections and an FP8 key-value cache. Nvidia provides serving recipes for vLLM and TensorRT-LLM, including configurations for one H100, eight H100s, one GB200 and one DGX Spark.
For faster generation, Lightning 3.5 includes three options:
- DSpark, a semi-autoregressive drafter recommended for DGX Spark and low-concurrency data-center deployments.
- DFlash, which uses a lightweight block-diffusion model to generate a draft block in one forward pass.
- Multi-Token Prediction, which trains the model to predict several future tokens at each position.
Nvidia recommends DSpark for DGX Spark deployments. The published vLLM configuration uses the nightly v0.27.1 image and can serve the full 1-million-token context window on H100 and GB200 configurations, subject to available memory.
Nvidia’s benchmark results
Nvidia evaluated the BF16 and NVFP4 versions using its NeMo Gym and NeMo Evaluator SDK harnesses. The company cautions that results may differ from vendor-reported scores. Selected results include:
The release page identifies the model version as GA, dated August 11, 2026. Nvidia does not provide independent verification of these benchmark results in the supplied release material.