What does JevStyle do?
Give a model text or JSON and typed yes/no, choice, or ordered-score questions. It scores the options in one forward pass. Use it for support ticket classification, request routing, or an additional review signal for agent tool calls. It does not generate free-form chat replies.
The Python package includes a local server, Python client, MCP tools, and agent skills. The 2B v4 PyTorch build also supports images; MLX and GGUF builds are text-only.
Models, package, and hosted services
Jev-Style names the model family, jev-style is the Python package, and JevStyle is the project and website. The model comparison separates releases, inputs, backends, and weight sizes.
After dependencies and weights are downloaded, local inference runs on your machine. The Hugging Face demo and hosted API are separate services that receive inputs sent to them. The hosted API currently uses 2B v4 Q4_K_M for text only, with up to 8,000 input tokens and one request at a time. Its limits differ from the local 2B v4 PyTorch model, which also supports images. See the current API reference.
Maintainer and official sources
JevStyle is an independent project maintained by Chaoliang Yan. Use this website, the source repository, the package registry, and the model cards to verify releases and capabilities.
License, provenance, and limits
The project code and published model weights are licensed under Apache-2.0. Each model card records its base model, training sources, and limitations. The project is not affiliated with or endorsed by TypeSafe AI (Jev), the Laya authors, Alibaba Cloud, or the Qwen team.
Training provenance differs by release. The 2B v4 model card discloses soft-label distillation from TypeSafe Jev 1.14 through its public API. Consult the specific model card instead of assuming that every release has the same sources.
Published performance numbers are project self-runs under stated protocols, not official leaderboard positions or guarantees for arbitrary tasks. Read the benchmarks and reproduction notes and evaluate accuracy, calibration, and fallback behavior on your own data.