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

Open Spark Jev

spark-s1: local Jev-inspired decision models on Qwen3.5-4B for NVIDIA DGX Spark, trained on code- and solver-labelled decision rows.

Modelresearch
View on GitHub ↗

Overview

Open Spark Jev, part of the Nokast community, is an independent local decision model inspired by Jev and System One, built for NVIDIA DGX Spark. The released model, spark-s1, takes a state and typed questions and returns structured decisions with probabilities and confidence, supporting Choice, Score and Noul for routing, urgency, tool-call safety and continue/stop/ask decisions. The state is encoded once and each question adds a short extra pass.

Version 6 trains Qwen3.5-4B with LoRA on 11,792 decision rows across 49 Jev-style task packs, labelled by code, policy engines and solvers rather than an LLM judge. The README reports 0.929 on its own held-out set, 83.1 on JevBench Intelligence and 74.9 ms p50 on vLLM in bf16, with an NVFP4 variant for better agent coverage and calibration. It is candid about being early-stage, not at parity with Jev, and the list summary notes its evaluation protocol records a past test-set leak.

What the decision model decides

TypeQuestionAnswers
ChoiceShould the agent continue, stop or ask?continue · stop · ask
Yes / noIs this tool call safe?true · false
ScoreHow urgent is this?—

How it works

  1. Deploy spark-s1 with vLLM on a DGX Spark.
  2. Send state plus typed questions.
  3. Receive decisions, distributions and confidence.
  4. Apply thresholds in your application.

Strengths

  • Fully local on DGX Spark.
  • Labels from deterministic sources rather than an LLM judge.
  • Documents known failure modes and leakage history.

Caveats

  • Early-stage; small per-pack test sets.
  • Not at parity with Jev.
  • Targets DGX Spark hardware.

Getting started

Download abhishek085/spark-s1-4b-v6 from Hugging Face and serve it with vLLM.

At a glance

Type
Model
Status
research
Language
Python
Runtime
vLLM on NVIDIA DGX Spark
Has code
yes

Use cases and integrations

How curated lists describe it

  • Open-source, local decision models inspired by TypeSafe.

    wh000wh000/awesome-jev-live

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

    hellogumbo/awesome-jev

  • A local decision model for NVIDIA DGX Spark, labelled from policy engines and solvers rather than an LLM judge. Its evaluation protocol records the time its own corpus leaked most of the test set into training.

    MrJev/awesome-jev

Frequently asked questions

What is spark-s1?

The decision model released by Open Spark Jev, built on Qwen3.5-4B with LoRA.

How is training data labelled?

By code, policy engines and solvers rather than an LLM judge.

Is it affiliated with TypeSafe?

No, it is an independent project inspired by Jev.

Related projects

  • Strands Decider 2B★ 612

    AWS Strands Labs' 2B decision model: Qwen3.5-2B with a pointer head scoring options and calibrated confidence, plus a CLI.

    ModelbetaPython
  • jev-style

    Open 0.8B and 2B Qwen3.5 decision models with a local systemone-compatible server, MCP tools, agent skills and a Claude Code guard.

    ServicebetaPython
  • brier

    Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.

    ProjectproductionPython
  • blink★ 22

    Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.

    ServicebetaC
  • Foq★ 11

    Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.

    ProjectbetaPython
  • browser-laya

    Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.

    ServicebetaTypeScript