Nemotron-Labs-Sigma-8B
AvailableNVIDIA research continuous diffusion language model, released on Hugging Face Nov 12 2026 as a base model plus an instruct (SFT) variant, under the NVIDIA Nemotron Open Model License (commercial use permitted). 8B parameters. Instead of generating one token at a time it denoises Gaussian noise in a learned 16-dimensional continuous token-embedding space and decodes text in parallel blocks (DDPM/DDIM samplers with self-conditioning), reusing committed tokens through a KV cache; backbone is a Ministral-style decoder-only Transformer. NVIDIA reports reasoning accuracy competitive with masked diffusion LMs and a 3.46x average self-speculative decoding speedup over autoregressive decoding on GB200 (vendor-reported). Distilled few-step checkpoints also published. Context window not stated on the model card.
Specifications
- License
- Open weights Β· NVIDIA Nemotron Open Model License
- Weights
- Downloadable
- Architecture
- Dense
- Parameters
- 8B
- Context window
- β tokens
- Max output
- β
- Knowledge cutoff
- β
- Price (in / out, $/M)
- β
- Modalities
- TextCode
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.