Why do weights load elsewhere but give no scores?

Use the bundled runtime files at pinned revisions through jev-style serve. Stock llama.cpp, Ollama, LM Studio, or raw generate paths load weights but cannot produce decision scores.

Are the benchmark numbers real or board-ranked?

Self-run on 231 public JevBench items: 2B at 73.6%, 9.5 above the 0.8B and top among its family, but inside error bounds versus decider-2b, with 42 systems higher and hosted Jev well ahead at 86.6%.

What latency and context limits should I expect locally?

Short requests take about 0.15-0.2s on the 0.8B with MLX on M1 Max after warmup, with CUDA graphs cutting medians sharply on RTX 5090. Long inputs up to 25,600 tokens are accepted or rejected, never silently cut.

How do I escalate only the unsure questions?

Cascades send only low-confidence questions from the small model to a larger one, including the named 2B plus JevK5-9B path. Eval reports accuracy, calibration, and automatable share at 1, 5, and 10% error budgets.

What are Jev-Style's current releases and their sizes?

Jev-Style-2B-Decision-v3 (1.27 GB at 4-bit) and Jev-Style-0.8B-Decision-v3 (0.53 GB at 4-bit, the default), both built on Qwen3.5 and runnable with no GPU or API key.

What question types and limits does the local Jev-Style server accept?

Three question types in one request: noul, choice with up to 255 options and score with 2-10 levels; inputs up to 25,600 tokens per call, and anything longer is rejected with an error rather than truncated.

What does CUDA graphs change in Jev-Style 0.4.0?

For the PyTorch backend on NVIDIA GPUs it is on by default; on an RTX 5090 median latency fell from 86.2 to 13.9 ms for the 2B and 41.6 to 11.3 ms for the 0.8B, with no top-1 answer changed across 4,992 questions.

CommunityJev Demos & Experiments

Jev Style

Local Qwen3.5 decision models in 0.8B and 2B sizes with fitted temperatures, systemone-shaped server, playground, guard, MCP tools, and eval harness.

See highlights

Open github.com

Stars: 16
Language: Python
License: APACHE-2.0
Published by: Community
Author: lawrence3699
Last verified: Oct 2026

lawrence3699/jev-style

User repository on GitHub

View on GitHub

Small, calibrated decision models on your own machine: systemone-compatible local server, 6 agent skills, Claude Code guard, MCP tools. Weights on Hugging Face.

Forks: 2
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Small, calibrated decision models you run on your own machine, plus the tooling to put them to work in AI agents.

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

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Ranking Jev for Moderation Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

One forward pass answers yes-no, up-to-255 choice, and 2-10 level score questions with calibrated probabilities.
Serve picks MLX, PyTorch, or optional llama.cpp with playground, demos, six skills, guard hook, and MCP tools.
Torch, MLX, and GGUF builds ship sized weights with matching runtimes for reproducible local scoring.
Eval measures your labels for accuracy, calibration, and safe automation share instead of trusting public boards.

Watch out

Treat both releases as small local options: hosted Jev leads on accuracy, the 2B lead over decider-2b sits inside confidence bounds, and stock runtimes cannot emit scores. 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

robodiag-harness

Terminal harness that checks ROS 2 robots with deterministic health tests, capability descriptions, SQLite history, and optional Jev evidence routing.