3B / 8B / 14B dense LM family; VLM-8B is listed as 9B on Hugging Face
262,144 tokens
NVIDIA Nemotron Open Model License (LM variants); NSCLv1 (VLM-8B)
Official open-weight Nemotron tri-mode family that unifies three decoding modes inside one model: standard autoregressive decoding, diffusion-based parallel decoding, and self-speculation where diffusion drafts and autoregressive decoding verifies against a shared KV cache. NVIDIA positions the family as a dense 3B/8B/14B release spanning base, instruct, and vision-language variants, designed to adapt to different deployment regimes rather than forcing a single decoding strategy. The technical report says the models are trained with a joint autoregressive-plus-diffusion objective and a two-stage pipeline that first strengthens left-to-right priors and then enables joint AR and diffusion training; the VLM extends the same backbone with a Pixtral-style vision encoder. Official release claims emphasize both quality and decode efficiency versus open autoregressive and diffusion baselines, including roughly 6x more tokens per forward than Qwen3-8B at comparable accuracy for the 8B model, about 4x higher throughput on SPEED-Bench with SGLang on GB200, and stronger self-speculation acceptance and speed than Qwen3-8B-Eagle3. NVIDIA also highlights practical deployment support through Transformers, vLLM, SGLang, and Docker, plus an optional LoRA-enhanced drafter for higher self-speculation acceptance lengths.