JevStyle Get API key

Local classification and routing for AI agents

Small open models that score your options in one call. Run them yourself, or use the free API.

downloads
124,135
likes
226
models
8
Get a free API key Try demo Run locally

One request, every option scored

A probability for each option

“I was billed twice. Please refund one charge.”

{"refund": {"type": "noul", "instructions": "The customer asks for money back."}}

Illustrative output: refund probability 0.967. Results depend on the model and input.

One state for all three questions: Hi, I was billed twice for my March plan. Please refund one of them today.

Run it yourself

CPU, CUDA and Apple silicon

pip install "jev-style[mlx]"    # Apple silicon
pip install "jev-style[torch]"  # CPU or CUDA
Quickstart

Use cases

All use cases

Benchmarks

Full evaluation

Jev-Style 2B v4: JevBench v1.4.1 public accuracy 74.9%; Decision Index 0.3 board: Public 44.37, Full 34.53.

JevBench is a single self-run. The Decision Index scores are quoted from the public board, where 2B v4 has been listed since 2026-11-04; Public is the suite anyone can run, Full adds the maintainers' private tests and is lower. Hardware, test protocols, confidence intervals, and limitations: read the model guide.

Start with your task

Browser agent: recorded runs and local setupLocal text classification in PythonLocal decision models on Apple siliconWorkflow routing with LLM fallbackCompare local routing toolsLocal receipt and screenshot decisionsReplace label-only LLM calls

Questions

Can JevStyle run offline?

Yes. After installing dependencies and downloading a model, inputs stay on your machine. The online demo and hosted API are separate services and receive what you send them.

Which model should I use?

Start with 0.8B v3 for text, then compare 2B v4 on your own examples. Image input needs the 2B v4 PyTorch build; MLX and GGUF builds are text-only.

Does it write chat replies?

No. It returns probabilities for yes/no, choice, and ordered-score questions. Send uncertain cases to a person or a larger model.

Project facts

Releases

The Python package is jev-style 0.6.3. 2B v4 was released on 2026-10-27: PyTorch supports text and images; MLX and GGUF builds are text-only. The default local text model is 0.8B v3.

Hosted API

The hosted API runs 2B v4 for text only, with an 8,192-token context and an 8,000-token input cap; there are no fixed per-account limits at the moment, capacity is shared, and a busy service returns 429 with Retry-After. Local models and the Hugging Face demo are separate entry points; hosted-service availability does not determine whether a local model works. Check dated service status and live health. Local installation needs no hosted API key.

Maintainer

Chaoliang Yan (GitHub: lawrence3699). JevStyle is independent and is not affiliated with or endorsed by TypeSafe.

Training provenance

2B v4 was trained in part on soft labels distilled from TypeSafe Jev 1.14 through its public API. JevStyle is independently maintained by Chaoliang Yan, with no TypeSafe affiliation or endorsement. The package does not bundle raw teacher-response datasets; this does not negate the use of teacher outputs in training. See the model and training sources.

Reviewed:

Start with one decision

Get a free API keyExplore models