Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

Nemotron Diffusion Decision Lab

Browser lab and TypeSafe-shaped API for asking typed questions of a Nemotron diffusion model and inspecting the distributions.

Modelresearch
View on GitHub ↗

Overview

Nemotron Diffusion Decision Lab packages a model, a browser explorer and a TypeSafe-shaped API in one container so you can ask Choice, Noul and Score questions about text or JSON and inspect probability distributions. The current branch defaults to the dense Nemotron-Labs-Diffusion-8B; earlier published images used a 14B model. The README is explicit that it is not Jev, not a TypeSafe service and not a calibrated replacement; jev-latest is only an input compatibility alias.

The explorer has samples, state and question editors, answer bars, raw JSON, local history and share links (which embed inputs, so avoid sensitive data). An H100-validated image runs the explorer and a classification-only vLLM API; with answer ordering and code reassignment, the README reports 174/231 on public JevBench. LoRA training experiments on 32K examples are documented. It requires a Hopper-class GPU.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich label applies?—
Yes / noProbability of yesyes · no
ScoreProbability per ordered level—

How it works

  1. Pull and run the container on an H100-class GPU.
  2. Enter state and typed questions in the explorer or API.
  3. Inspect per-option probabilities and confidence.
  4. Compare methods such as answer ordering.

Strengths

  • Self-hosted, TypeSafe-shaped API for experiments.
  • Honest disclaimers about identity and calibration.
  • Pinned model revision and immutable image digests.

Caveats

  • Needs an H100/H200-class GPU and CUDA 13 driver.
  • Not calibrated like Jev.
  • Share links include entered data.

Getting started

docker pull ghcr.io/pst2154/nemotron-jev:14b-vllm-ordered-20260922 and run it with the NVIDIA Container Toolkit.

At a glance

Type
Model
Status
research
Language
Python
Runtime
Docker, vLLM, browser explorer
Has code
yes

Use cases and integrations

How curated lists describe it

  • Browser lab for asking typed questions of a dense diffusion model and inspecting the distributions, with a disclaimer that it is neither Jev nor a TypeSafe service.

    MrJev/awesome-jev

Frequently asked questions

Is this Jev?

No; it uses a Nemotron diffusion model behind a TypeSafe-shaped interface.

What hardware does it need?

An H100 or H200-class Hopper GPU; H100 80 GB was tested.

How does it score on JevBench?

The README reports 174/231 for the ordered 14B image.

Related projects

  • TypeAR★ 946

    TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.

    ProjectbetaPython
  • jev-dllm

    Reads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.

    ProjectresearchPython
  • metask-jev

    Open Jev-class decision models on Qwen3.5 with candidate-logit readout; the 4B model reports 80.1% on JevBench's 231 items.

    ProjectresearchPython
  • system-one-open

    Open Jev replica on Gemma 4 E2B and Gemma 3 270M served on Modal, with eight demos and a TypeSafe eval head-to-head.

    ProjectresearchPython
  • jev-position-test

    Small experiment: reordering enum options changed a local Jev clone's answers on its old default, while hosted Jev's did not.

    DemoresearchPython
  • JevK5★ 150

    Apache-2.0 open-weight Jev alternative that answers yes/no, choice and score questions with per-option probabilities in one pass.

    LibrarybetaPython