---
type: "post"
id: "nvidia-publishes-nemotron-labs-sigma-distill-3b-nfe4-8e6c98"
slug: "nvidia-publishes-nemotron-labs-sigma-distill-3b-nfe4-8e6c98"
title: "NVIDIA publishes Nemotron-Labs-Sigma-distill-3B-nfe4 weights"
description: "NVIDIA published Nemotron-Labs-Sigma-distill-3B-nfe4, an open-weight, three-billion-parameter continuous-diffusion language model distilled for four-step generation."
retrieval_nugget: "NVIDIA published Nemotron-Labs-Sigma-distill-3B-nfe4, an open-weight, three-billion-parameter continuous-diffusion language model distilled for four-step generation."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["open-models","inference"]
entities: ["NVIDIA","Nemotron-Labs-Sigma-distill-3B-nfe4","Nemotron-Labs-Sigma","Hugging Face"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "NVIDIA has published weights for a new 3B Nemotron-Labs-Sigma distilled model, distinct from its previously covered Nemotron releases."
prior_refs: ["https://newruntime.com/posts/nvidia-releases-open-weight-nemotron-3-5-transcribe-3d2f81/"]
source_url: "https://huggingface.co/nvidia/Nemotron-Labs-Sigma-distill-3B-nfe4"
source_urls: ["https://huggingface.co/nvidia/Nemotron-Labs-Sigma-distill-3B-nfe4"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:nvidia-publishes-nemotron-labs-sigma-distill-3b-nfe4-8e6c98"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/nvidia-publishes-nemotron-labs-sigma-distill-3b-nfe4-8e6c98/","markdown":"https://newruntime.com/posts/nvidia-publishes-nemotron-labs-sigma-distill-3b-nfe4-8e6c98.md","json":"https://newruntime.com/posts/nvidia-publishes-nemotron-labs-sigma-distill-3b-nfe4-8e6c98.json"}
markdown_url: "https://newruntime.com/posts/nvidia-publishes-nemotron-labs-sigma-distill-3b-nfe4-8e6c98.md"
---

# NVIDIA publishes Nemotron-Labs-Sigma-distill-3B-nfe4 weights

## Retrieval answer

NVIDIA published Nemotron-Labs-Sigma-distill-3B-nfe4, an open-weight, three-billion-parameter continuous-diffusion language model distilled for four-step generation.

## Summary
NVIDIA published Nemotron-Labs-Sigma-distill-3B-nfe4, an open-weight, three-billion-parameter continuous-diffusion language model distilled for four-step generation. Its fixed denoising schedule and local NVIDIA-GPU requirements make it a concrete artifact for teams testing diffusion decoding, not a drop-in general-purpose inference option.

## What happened
NVIDIA’s Hugging Face model card describes Nemotron-Labs-Sigma-distill-3B-nfe4 as a student model produced through Phased Distribution Matching Distillation from a multiscale SUSReg-regularized 3B teacher in the Nemotron-Labs-Sigma family. The model card identifies the base lineage as the same as the released Nemotron-Labs-Sigma-3B model. The checkpoint is version 1.0 and uses a Ministral-style decoder-only Transformer with grouped-query attention, SwiGLU, rotary position embeddings, 26 layers and a 3,072-wide hidden state. A continuous-diffusion generation head operates in a 16-dimensional continuous token-embedding space; an auxiliary autoregressive readout head is preserved.

The defining operating constraint is NFE=4. The student generates a 32-token block through exactly four denoising function evaluations followed by one final readout. NVIDIA says the bundled inference code does not allow an arbitrary step override: this variant was trained for four evaluations, and other step counts would use untrained phase endpoints. The card says sibling variants span NFE 2, 4, 8 and 16 for each model size. Generation can produce up to 4,096 tokens total from a maximum 4,096-token context, including the prompt. Stop strings are checked after each completed block; generation also stops at the caller’s token budget. There is no token-ID EOS stopping.

The model accepts plain-text input and supports English. NVIDIA recommends few-shot or benchmark-style prompting. It requires PyTorch 2.4 or later, Transformers 4.55 or later, and a CUDA-capable NVIDIA GPU running bfloat16 on Linux. The card lists compatibility with Ampere, Blackwell, Hopper and Lovelace. CPU inference is not supported: NVIDIA says the diffusion sampler is numerically degraded on CPU and produces unusable output.

The repository includes Transformers examples using `trust_remote_code=True`, and serving examples for vLLM and SGLang through OpenAI-compatible endpoints; Docker Model Runner is also named. The model runs offline on a user’s own hardware, and NVIDIA does not host a service for it. The model is released under the NVIDIA Nemotron Open Model License for commercial or non-commercial use. NVIDIA reports mean accuracy over 10 generation seeds: 82.23% on GSM8K (CoT), 59.13% on MATH (Minerva), 52.13% on HumanEval, 48.72% on HumanEval+, 40.62% on MBPP and 58.81% on MBPP+, with a six-task mean of 56.94%. The card gives two-standard-error ranges for each result.

## What the archive adds
New Runtime’s earlier coverage described NVIDIA’s open-weight Nemotron 3.5 Transcribe model and Granary v2 speech dataset on [November 6, 2026](https://newruntime.com/posts/nvidia-releases-open-weight-nemotron-3-5-transcribe-3d2f81/). This is a new model release, not a new fact about that speech-model story; the shared element is NVIDIA’s publication of open model artifacts.

## Trend signal
This release is consistent with the move toward evaluating agent and model economics by completed work rather than model calls alone, a theme in [Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/). The model card specifies a fixed number of denoiser evaluations per generated block, but does not report end-to-end cost or verified task completion. It therefore offers a measurable inference mechanism, not evidence that the economic unit has shifted in practice.

## Why now
The Hugging Face model card identifies the release date as November 12, 2026. It does not explain the timing beyond publishing this checkpoint within the Nemotron-Labs-Sigma family, which also includes other model sizes and inference-step variants.

## Why it matters
For teams investigating alternatives to token-by-token autoregressive decoding, the release makes one trade-off unusually explicit: this checkpoint’s generation schedule is fixed at four denoising calls for each 32-token block, with one additional readout. That can make a controlled experiment easier to define, but the number of calls alone does not establish faster or cheaper serving. Teams still need to measure latency, throughput, GPU memory, output quality and serving compatibility under their own prompts and workloads.

The deployment boundary also matters. The checkpoint is downloadable and usable under NVIDIA’s model license, but it is not an inference-provider endpoint, and the card requires NVIDIA GPU hardware and custom code loading. Builders should treat the reported benchmark scores as NVIDIA’s model-card results rather than a direct comparison with a chosen production baseline. For a research or prototyping team, the artifact is useful precisely because its constraints are stated; for a production team, those constraints define the first validation questions.

## Limits and uncertainty
The benchmark table reports results over 10 generation seeds and identifies the metrics, but the model card does not establish comparative performance against autoregressive baselines under matched hardware, latency or cost conditions. It also does not report production throughput, GPU memory use, or the cost of serving a workload. The card’s operational claims, including unusable CPU output and compatible GPU generations, are NVIDIA’s statements and are not independently verified here. English is the only supported language stated. A model license permitting commercial use does not by itself resolve a team’s compliance, safety or quality requirements for a specific deployment.

## Skill to build now
If your team has a CUDA-capable NVIDIA GPU, reproduce the model-card setup on a small fixed prompt set before considering deployment. Keep the checkpoint’s NFE=4 setting unchanged, record latency and peak GPU memory per generated block, and score outputs against the same task-specific tests you use for an autoregressive baseline. This isolates whether the fixed-step diffusion path is useful for your workload without mistaking benchmark scores for a serving result.

## Reader takeaway
Nemotron-Labs-Sigma-distill-3B-nfe4 is an open-weight, English-only diffusion model with a precisely fixed four-evaluation generation path and a local NVIDIA-GPU requirement. It gives builders a concrete checkpoint to test, but its model-card benchmarks do not yet answer whether it improves real-world latency, cost or task quality for their applications.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
