NVIDIA releases Nemotron-Labs-Sigma diffusion language models
Understand this story
NVIDIA released the Nemotron-Labs-Sigma series of continuous diffusion language models, comprising 14 open-source checkpoints covering 3B and 8B base, instruct, few-step distilled, and self-speculative versions, with the models performing denoising in a 16-dimensional embedding space. All 14 checkpoints are released on Hugging Face under the NVIDIA Nemotron Open Model License, and the model cards state they can be used commercially. Running them requires a CUDA GPU supporting bfloat16 (Ampere, Lovelace, Hopper, or Blackwell); CPU inference is not supported.
AI overview based on source reports · Updated 1 hr ago
Coverage timeline
Follow the coverage to explore different perspectives.
- MarkTechPostNVIDIA AI Releases Nemotron-Labs-Sigma, a Continuous Diffusion Language Model with 14 checkpoints at 3B/8B
NVIDIA released Nemotron-Labs-Sigma, a family of 14 open checkpoints at 3B and 8B scale that denoises text in a learned 16-D continuous embedding space rather than on discrete tokens. Sigma-8B base averages 71.16 on NVIDIA's 6-task suite versus 70.22 for NLD-8B, and the self-speculative 8B checkpoint runs 3.46x faster than AR decoding on average on one GB200 with SGLang. Weak spots include 17.00 on AIME25, a 4,096-token base context (8,192 for Sigma-8B-instruct) and GPU-only inference in bfloat16.
Activity for this story
Not enough continuous observations to show a trend yet.