What NVIDIA shipped

Almost every language model in production writes the same way: one token, then the next, each step waiting on the last. Sigma does not. It treats a token as a point in a continuous 16-dimensional space, starts from noise, and refines a whole block of tokens at once, the way a diffusion image model refines a picture. Once a block settles, it is committed to a standard KV cache and the model moves on.

The release is a family, not a single model. At each of 3B and 8B there is a base model, an instruct variant, distilled versions rated for as few as two function evaluations per block, and a self-speculative version in which a diffusion head drafts a block and the model's own autoregressive head verifies it. Fourteen checkpoints in total, all on Hugging Face under the NVIDIA Nemotron Open Model License; the 8B card states the model is ready for commercial or non-commercial use.

This is research infrastructure more than a product. The base card describes a pretraining-quality base model for research on sampling, few-step distillation, trajectory steering and fine-tuning, not an assistant. The training corpus, about 1.3 trillion tokens, is internal and not released. What NVIDIA has actually given away is the architecture and the weights, plus a paper that has been through at least one major revision.

Where the speed numbers come from

The figure traveling with this release is "up to 4.2x faster than autoregressive." Read the cards and the paper coverage, and a more layered picture appears.

The base 8B model, running its documented sampler setup, is documented at roughly 120 to 140 tokens per second for 256 generated tokens on a single H100. The card is explicit that fewer denoising steps mean faster generation at some quality cost. That trade is what the distilled checkpoints are for: they are trained for a fixed, small number of evaluations per block, and the 4.2x figure from the paper belongs to them.

The self-speculative variant's 3.5x speedup, as reported from the paper, carries its conditions with it. It is measured before supervised fine-tuning, and the spec card's own wording measures speedups against the same model's autoregressive decoding, in SGLang in bfloat16 on one GB200, one request at a time. Neither condition makes the number false. Both make it a laboratory result rather than a deployment expectation, and no independent group has replicated any of these figures yet.

There is also a hardware wall. Sigma needs a CUDA-capable NVIDIA GPU running bfloat16. The card states that CPU inference is not supported because the sampler is numerically degraded on CPU, and no quantized variants are among the 14 checkpoints, so the usual route by which an open 8B model reaches laptops and small servers is not open here yet. Combined with a 4,096-token context and English-only support, the practical audience today is researchers with data-center cards, not developers looking for a faster local model.

The quality comparison, on NVIDIA's terms

On the six-task suite reported in the paper and relayed by MarkTechPost, Sigma-8B base averages 71.16, ahead of NLD-8B, a masked diffusion model, at 70.22. The comparison that matters here is against other diffusion models, not against the best conventional models of its size. The margin is under one point, on the maker's own suite, averaged over the maker's chosen tasks. It establishes that continuous diffusion can be trained at 8B scale and score in this range. It does not establish an advantage a buyer would notice.

What is genuinely new is what the continuous space makes possible downstream. Because the denoising path is a smooth trajectory, NVIDIA can distill it to very few steps and steer it, and the paper reports holding accuracy and diversity through that distillation for math and code sampling. If that holds up under outside testing, the interesting use is not raw chat speed. It is generating many candidate solutions in parallel for problems where a system samples, checks and keeps the best one.

What to watch

Three things will decide whether Sigma is a curiosity or a direction. First, independent benchmarks: every number in this article traces back to NVIDIA, and the weights have been public since November 12, so outside results should not take long. Second, serving support: the spec card's speedups were measured in SGLang, and real adoption follows serving frameworks, not model cards. Third, context and quantization: a 4,096-token ceiling and a bfloat16-only release are research-grade constraints, and both are fixable in a follow-up release if NVIDIA decides this family has a future beyond the lab.

Sources

  1. [1] NVIDIA, Nemotron-Labs-Sigma-8B model card, Hugging Face (release date November 12, 2026)Read source
  2. [2] NVIDIA, Nemotron-Labs-Sigma-distill-spec-3B model card, Hugging FaceRead source
  3. [3] Yang et al., "Large Language Continuous Diffusion Models", arXiv:2610.02665Read source
  4. [4] MarkTechPost, "NVIDIA AI Releases Nemotron-Labs-Sigma, a Continuous Diffusion Language Model with 14 checkpoints at 3B/8B", November 20, 2026Read source