Oossa

NVIDIA releases Nemotron‑Labs‑Sigma, a faster continuous diffusion language model

NVIDIA makes 14 open checkpoints of 3B‑ and 8B‑parameter models that decode up to 4.2× faster than traditional autoregressive LLMs.

By Published by Oossa: 1 min read

Museums Victoria · Unsplash

NVIDIA announced a new family of language models called Nemotron‑Labs‑Sigma. The company published 14 checkpoints – three‑billion‑parameter (3B) and eight‑billion‑parameter (8B) versions – on the Hugging Face model hub. The models use a continuous diffusion process that denoises text in a learned 16‑dimensional embedding space. NVIDIA says this lets the models generate text up to 4.20 times faster than standard autoregressive decoding.

What makes the diffusion approach different?

Traditional large language models (LLMs) generate one token after another, a method called autoregressive decoding. Nemotron‑Labs‑Sigma instead starts with a noisy representation of the whole sentence and iteratively removes noise, similar to how image diffusion models create pictures. Because the diffusion runs in a low‑dimensional space, NVIDIA can distill the process into very few steps – as few as two forward passes for the 3B model – while keeping quality high. The company also added a “speculative” one‑step mode for the 8B checkpoint that speeds up throughput another 3.5× on benchmark GB200.

Performance on benchmarks

In the paper that introduced the models, the authors compared Sigma to state‑of‑the‑art discrete diffusion models and to standard autoregressive baselines on a range of tasks. After pre‑training, Sigma beat the others on math and coding benchmarks such as GSM8K, HumanEval, and MBPP. After supervised fine‑tuning, the gains grew larger on harder problems like MATH‑500 and AIME, with the biggest improvements seen in code generation. The few‑step distilled versions delivered roughly 8× speed‑up on math tasks and 4× on coding tasks with only a small loss in accuracy.

Why it matters

For developers who need quick responses – such as code assistants or real‑time chatbots – Sigma’s faster decoding could lower latency and cloud costs. The open checkpoints also let researchers experiment with diffusion‑based generation without building the models from scratch. Independent verification of speed claims is still pending, so real‑world gains may vary.

Was this article useful?
Share

Read next

Oossa · Newsletter

The week in AI, explained

Every Monday: the stories worth knowing, in plain language. Free, no spam.