Unknown · od 1. Oktober 2026 (Kev 1.0 mit Kev-27B/Kev-9B-Update; erste Version Kev-0.5B bereits am 18. September 2026) · 2× · naposledy 03. 10. 2026
Kev is an open family of four "decision models" (0.8B to 27B parameters) released by Jared Palmer, built on Qwen3.5/Qwen3.8 base models, positioned as an open-source alternative to TypeSafe's hosted "Jev" model. The models answer typed questions (yes/no, multiple-choice, rating) in a single forward pass with calibrated probabilities instead of standard text generation. Kev runs locally via CUDA/ROCm or on Apple Silicon Macs through MLX and exposes an API compatible with TypeSafe's System One contract (/v1/systemone), so existing TypeSafe SDK clients work unchanged. The project is licensed under Apache-2.0 and distributed via GitHub and Hugging Face as weights/code.
Vlastnosti
| Throughput/Latency | Kev-4B: 721 ms for 5 questions (new text) / 136 ms when cached, on 32GB M5; Kev-0.8B: 149 ms for new text |
| License | Apache-2.0 (open source, commercial use allowed) |
| Platform | CUDA/ROCm GPUs and Apple Silicon Macs via MLX; runs locally or as HTTPS endpoint via Modal |
| Price | Free / open-source (self-hosted); via OpenRouter e.g. Kev-4B from $0.042/M input tokens, $0/M output tokens |
| Protocol Compatibility | Compatible with TypeSafe's System One API (POST /v1/systemone); TypeSafe Python SDK works unchanged against a Kev server |
| Release Date | Kev-0.5B: September 18, 2026; Kev 1.0 (0.8B–27B family): October 1, 2026 |
| Supported Models/Providers | Kev-0.8B, Kev-4B, Kev-9B (Qwen3.5 base), Kev-27B (Qwen3.8-27B, post-trained) |