Should I swap this in wherever I currently call Jev?

For local control and latency experiments, not as a drop-in Jev replacement. The author states it trains on far less data, is not at parity, and varies substantially across task families.

Which should I run: v6 or the newer v8 build?

Pick v6 for general accuracy and hard reasoning, and v8 for agent-harness coverage and calibration. V8 adds tool choice, finish, command risk, and output screening with fewer confident wrong answers.

What hardware and setup does local serving actually need?

Clone, install the inference extras, pull the checkpoint once, then serve locally and call the gate endpoint or the Python helpers. A GPU is needed for the reported speeds, tested on DGX Spark.

How do the two smaller control models compare?

Control-core favors accuracy at higher latency while the smaller embedding-style build favors speed with weaker relevance. Both trail the prompted baseline on the demo harness, so test on your own decision sites.

What are the actual numbers behind those choices?

spark-s1-4b-v6 reports 0.929 on its own held-out split and 83.1 on JevBench, at a p50 of 74.9 ms in bf16 on vLLM. The NVFP4 build keeps the same reported accuracy, scores 83.2 on JevBench and runs at 53.3 ms, about 1.40x faster. v8 trades a little of that for agent decisions, with JevBench hard at 0.586 against v6-nvfp4's 0.622. All of these are the project's own measurements.

How strict is the default gate, and can I loosen it?

Strict by default. The gate endpoint returns a choice with probabilities, a confidence, a policy_action and a policy_trace, and allow needs 99.5% confidence or the agent asks a human instead. You can lower that per request by sending a policy object with auto_allow_threshold set to 0.9, so the bar is a parameter rather than a constant.

Are the latency figures something you can reproduce?

Not from the same code path. The README notes the sample output is HF Transformers in-process on a single request, while the table's p50 figures are vLLM, which is faster and recommended for production. The speeds also assume an NVIDIA DGX Spark; other GPUs should work but are untested, and the v8 build's latency is not re-measured at all.

CommunityJev Demos & Experiments

Open Spark Jev

Independent open-source local decision model inspired by Jev, with a spark-s1 family plus smaller control models for agent gates, served locally with probabilities and policy actions.

See highlights

Open github.com

Stars: 22
Language: Python
License: APACHE-2.0
Published by: Community
Author: abhishek085
Last verified: Oct 2026

abhishek085/open-spark-jev

User repository on GitHub

View on GitHub

Open-source, local decision models inspired by TypeSafe’s Jev and System One - built on Qwen3 for NVIDIA DGX Spark.

Forks: 4
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Part of Nokast , an open-source AI community. An open-source, local implementation inspired by TypeSafe's Jev and System One models, built for NVIDIA DGX Spark. The model is called spark-s1 .

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Evaluation Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Builds spark-s1 checkpoints that answer Choice, Score, and Noul questions with distributions and confidence.
Uses a restricted first-token readout rather than generating text, with state encoded once per decision.
Ships a local gate API with policy actions where allow needs very high confidence unless you lower it.
Adds smaller control builds for guardrail, routing, ranking, and triage sites inside an agent harness.

Watch out

This is an early-stage independent project with known failure modes and small per-pack tests; it is not at parity with TypeSafe Jev. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

open-jev

Open harness that renders allowed JSON answers on a diffusion canvas, denoises for a fixed step budget, then picks the top allowed tokens from final logits with no training.