Updates

Documentation-only release: clarify official local inference, TypeSafe teacher-label provenance, and separate hosted-service status. Runtime behavior, dependencies, model pins and weights are unchanged.

Refreshed the PyPI description and added direct links to model selection and benchmark protocols. Runtime behaviour is unchanged.

Added runnable support-ticket, fallback-routing and receipt-review recipes, plus a Semantic Router example.

The first Jev-Style model that reads images, through its PyTorch build. MLX and GGUF builds are text-only.

Added 2B v4 and image requests. --release 2b now selects 2B v4; the default release stays 0.8B v3.

Added compressed request bodies and tighter request checks for compatibility with current official SDKs.

A cascade configuration: our 2B answers first and sends uncertain questions to a larger model.

Added CUDA graphs for the PyTorch backend and confidence cascades across model tiers.

Released with PyTorch, MLX and GGUF builds, and available through jev-style 0.3.0.

The package became installable with pip and gained an eval command for comparing engines on your own questions.

Shipped a local API server, six agent skills, a Claude Code guard, an MCP server and an evaluation command.

A smaller long-input, multilingual model, and the default local text model today.

A 322M decision model for local Mac agents.

Calibrated English classification at 2B.

The first published decision model. Superseded by v3 and v4.