NVIDIA Nemotron-Labs-Diffusion: New Tri-Mode LLM

NVIDIA researchers have released Nemotron-Labs-Diffusion, a new language model family. This model unifies three different decoding modes within a single architecture. It supports autoregressive, diffusion-based parallel, and self-speculation decoding. The model comes in 3B, 8B, and 14B parameter sizes. It also includes base, instruct, and vision-language variants. Standard language models generate text one token at a time, which limits their speed. Diffusion models generate multiple tokens in parallel, but have often lagged in accuracy. Nemotron-Labs-Diffusion is trained on a joint objective, allowing it to operate in three modes depending on how it's used. The AR mode is good for cloud serving. Diffusion mode denoises multiple tokens in parallel, allowing for faster processing. Self-speculation mode drafts candidate tokens and then verifies them, all within the same model. This new approach could significantly improve how language models generate text and utilize hardware.