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.
Overview
Jev-Style is a family of small decision models built on Qwen3.5, plus tooling to use them in agents. The default Jev-Style-0.8B-Decision-v3 is 0.53 GB at 4-bit and the 2B release is 1.27 GB. Given text or JSON and typed questions, a model returns a calibrated probability for every option of every Noul, Choice (up to 255 options) and Score (2 to 10 levels) question in one forward pass, over inputs up to 25,600 tokens. The local server follows the public systemone request shape, so Jev-compatible clients can call a laptop instead.
The repo provides `jev-style serve` (MLX on Apple silicon, PyTorch on CUDA or CPU, optional llama.cpp), six agent skills, an MCP server with decide, noul, choice and score tools, a Claude Code PreToolUse guard that maps four Noul checks and a risk Score to allow, ask or deny, and an eval command that reports how many decisions can be automated at a given error budget. The README is candid that hosted Jev is more accurate: 86.6% on JevBench public items versus 73.6% for the 2B and 64.1% for the 0.8B.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Guard checks on a pending tool call (four yes/no risk checks) | — |
| Score | How risky is this tool call? | — |
| Other | Any user-defined Noul, Choice or Score question served locally | — |
How it works
- pip install and run jev-style serve to host a local /v1/systemone API
- Send state and typed questions from SDKs, skills or MCP tools
- Model returns calibrated probabilities in one pass
- Code-owned thresholds act; the guard falls back to ask if the model call fails
Strengths
- Runs locally with no API key, GPU or training; data stays on the machine
- README reports 0.15-0.2 s per short request on an M1 Max with MLX, and 11-14 ms on an RTX 5090 with CUDA graphs
- Fitted temperatures for calibrated probabilities
- Cascades send only uncertain questions to a larger model
Caveats
- Less accurate than hosted Jev on every published comparison per the README
- 2B lead over similar open models is within the confidence interval
- Inputs over 25,600 tokens are rejected
Getting started
pip install "jev-style[torch]" (or "jev-style[mlx]" on Apple silicon), then `jev-style serve`.At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- Local server / MCP server / Claude Code hook
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Local runtime: `pip install "jev-style[torch]"` (or `[mlx]` on Apple silicon) serves an open 0.8B Qwen3.5 decision model behind a System One-compatible `/v1/systemone` API that answers `Choice`, `Score`, and `Noul` questions with calibrated probabilities in one pass over inputs up to 25,600 tokens (0.15–0.2 s per short request with MLX on an M1 Max, after the first call), and ships a Claude Code guard that turns four `Noul` checks and a risk `Score` into allow, ask, or deny through code-owned thresholds.
yibie/awesome-jev0.8B Qwen3.5 decision model you install with `pip install "jev-style[torch]"` (or `[mlx]` on Apple silicon) and run locally (PyTorch, MLX, or llama.cpp with a separately built scorer) behind a `/v1/systemone`-compatible server: Choice, Score, and Noul in one pass, plus an MCP server and a Claude Code guard hook
AnotiaWang/awesome-decision-modelsOpen 0.8B Qwen3.5 decision model installed with `pip install "jev-style[torch]"` (or `[mlx]` on Apple silicon) that answers Choice, Score and Noul questions locally behind a `/v1/systemone`-compatible server, while code owns the thresholds (its Claude Code guard maps four Noul checks and a risk Score to allow, ask or deny, and falls back to ask when the model call fails); its README reports hosted Jev as more accurate on every published comparison set.
cobanov/awesome-jevSmall Qwen3.5 decision models at 0.5 and 1.3 GB in 4-bit, with a systemone-compatible local server, agent skills and a Claude Code guard.
MrJev/awesome-jev
Frequently asked questions
Is Jev-Style as accurate as hosted Jev?
No. The README reports hosted Jev 1.13 at 86.6% on JevBench public items versus 73.6% (2B) and 64.1% (0.8B).
Can existing Jev clients use it?
Yes, the server follows the public systemone request shape, so compatible clients can point at it.
Does it need a TypeSafe key?
No. It runs locally without an API key or GPU.
Related projects
Runs Laya System 1 models on a Mac's GPU in Rust with custom Metal kernels, serving the same /v1/systemone API as laya serve.
ModelbetaRustRust engine that answers noul, choice and score from any LLM in one prefill via logprobs, wire-compatible with /v1/systemone and MCP.
ModelbetaRustMCP server exposing Jev-style decision models served locally by Ollama's /v1/systemone as classify, yes/no and score tools.
MCP serverbetaTypeScript- 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 spark-s1: local Jev-inspired decision models on Qwen3.5-4B for NVIDIA DGX Spark, trained on code- and solver-labelled decision rows.
ModelresearchPythonDocker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.
MCP serverbetaPython