Tasks and inputs
Nemotron-Labs Sigma 8B is listed for Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Nemotron-Labs Sigma 8B is NVIDIA's open continuous diffusion language model, denoising 64-token blocks in parallel; English, GPU-only, 8K context.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights under the NVIDIA Nemotron Open Model License; not offered by any hosted provider.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Most language models write left to right, one token at a time. Nemotron-Labs Sigma 8B does not. It is a continuous diffusion language model: it starts each block of text as Gaussian noise in a 16-dimensional embedding space and denoises the whole block in parallel over a set number of steps, then commits it and moves on.
NVIDIA lists the instruct release as 12 November 2026, though the repositories were created in early October and the distilled variants landed on 9 November. It accompanies the paper "Large Language Continuous Diffusion Models".
Parallel block generation changes the speed equation. On one H100 in bfloat16, NVIDIA reports about 120-140 tokens per second for 256 generated tokens at 32 denoising steps per 64-token block. Fewer steps run faster, at a quality cost, which is what the distilled variants are for.
This is a research-grade architecture with a production-friendly licence. The commercial diffusion models you can call today, such as Mercury 2.5, are closed; this one lets you look inside.
Quality drops noticeably below about 16 denoising steps for the pretrained models. If you need 4 or 8 steps, use a distilled checkpoint instead of turning down the dial on this one.
| Steps | Guidance | Score temperature |
|---|---|---|
| 32 | 3 | 0.3 |
| 16 | 3 | 0.4 |
| 8 | 3 | 0.6 |
| 4 | 2 | 0.8 |
Use DDIM on the log-SNR grid with 32-token blocks, per NVIDIA; the default block is 64.
trust_remote_code=True (PyTorch 2.4+, Transformers 4.55+).Training totalled about 1.3T tokens: roughly 1T of autoregressive initialisation, then about 0.3T of diffusion training, then SFT. The licence is the NVIDIA Nemotron Open Model License, which differs from the OpenMDW terms on NVIDIA's other October releases.
Researchers studying non-autoregressive generation. Teams prototyping latency-critical short completions. Anyone curious whether diffusion decoding suits their workload. Skip it for long documents or multilingual work.
To evaluate: fix a set of short prompts, generate at 32, 16 and 8 steps with the base instruct model and the matching distilled variant, and plot quality (human or task score) against tokens per second. Use the bundled chat template, except for raw-prompt code benchmarks.
More on the NVIDIA provider page and under text generation.
Compare the model on the work you need to complete.
Nemotron-Labs Sigma 8B is listed for Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 8.19K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under NVIDIA Nemotron Open Model License. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Tencent
Tencent Hy4 Flash is the cost-focused companion to Hy4, keeping vision and agent skills at a far lower token price. Availability is still unsettled.
Alibaba
Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.
Sarvam
Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.
IBM
Granite 4.1 8B LLMON Preview is IBM's model trained on block-structured LLMON prompts to resist indirect prompt injection in agents and RAG.
Alibaba
Qwen4 Max Preview was Alibaba's five-day early-access build of its Qwen4 flagship, offered only through a Token Plan subscription before GA qwen4-max.
speakleash
Bielik-11B-v3.1-Instruct is SpeakLeash's open Polish-first 11B instruct model, trained on ~56 languages, with five reasoning levels and tool calling.