NVIDIA's Sigma collection includes 14 diffusion checkpoints
MarkTechPost reports that NVIDIA released 14 open Nemotron-Labs-Sigma checkpoints at 3B and 8B parameter scales. The evidence does not establish when the full collection became available: Hugging Face lists Sigma-8B with a date of 11/12/2026.
Key takeaways · 4
- 01
The reported collection has 14 open checkpoints at 3B and 8B parameter scales.
- 02
Eight checkpoints are few-step distilled models: distill-3B and distill-8B at NFE 2, 4, 8, and 16.
- 03
NVIDIA reports a 3.46-times average throughput speedup for the 8B self-speculative model over autoregressive decoding on a GB200.
- 04
Plan for NVIDIA GPU hardware: the models require CUDA and bfloat16, and CPU inference is unsupported.
A block-based diffusion approach
Sigma is described by NVIDIA as a family of continuous diffusion language models aimed at reasoning and coding.[3] Rather than generate text only through the conventional autoregressive approach, Sigma generates one block at a time, denoising a compact representation conditioned on preceding context before reading out tokens.[3] MarkTechPost reports that this denoising takes place in a 16-dimensional embedding space.[1] NVIDIA says the smooth ODE trajectory enables inference-time steering as well as few-step generation.[3] For teams assessing model architecture, these are the design features to compare with their existing generation workloads.
Checkpoint options and reported results
The 14-checkpoint collection includes eight few-step distilled versions: distill-3B and distill-8B at NFE 2, 4, 8, and 16.[1] NVIDIA says the distilled models generate eight tokens per forward pass on math tasks and four on coding tasks, with what it characterizes as negligible accuracy loss.[3] On NVIDIA's six-task benchmark suite, Sigma-8B base averaged 71.16, compared with 70.22 for NLD-8B in diffusion mode.[1] Results vary by evaluation: Sigma-8B-instruct's 10-task average was 59.14, below NLD-8B's 59.92.[1] Benchmark figures should therefore be read in their stated task and model context.
Deployment and timing caveats
NVIDIA says the models are ready for commercial use.[1] Hugging Face lists a requirement for a CUDA-capable NVIDIA GPU running bfloat16 and says CPU inference is unsupported.[2] The models run offline on a user's own hardware, with no service hosted by NVIDIA.[2] Hugging Face reports generation throughput of roughly 120 to 140 tokens per second on a single H100 GPU.[2] That figure is tied to the stated hardware, so teams should validate performance against their own setup. The listed Sigma-8B date is 11/12/2026; it does not by itself establish when the complete 14-checkpoint collection became available.[2]
Sigma gives engineering and infrastructure teams another model family to evaluate for reasoning and coding, with multiple distilled checkpoints and a reported self-speculative speedup. The practical decision depends on whether the team's hardware, task-specific evaluation results and deployment preferences fit the model's requirements.
Why it matters
AI training for your team
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
How this developed
20 November 2026
NVIDIA's Sigma collection includes 14 diffusion checkpoints
