Kev ships 0.8B, 4B and 9B open-weight decision models that drop into TypeSafe's API, 1,684 stars in four days.

Jared Palmer published Kev September 17 on Qwen3.5, Apache-2.0 with training code and frozen eval suites. One request carries yes/no, multiple-choice and rating questions that share input text but can't read each other, and the API matches TypeSafe's System One so their Python SDK points at a local server. The Qwen3.5 generation on September 20 cost about $95 of rented Modal H100 time plus $0.03 in API calls, and Palmer published the benchmark he failed: Kev-9B at 0.837 on the locked test set against 0.780 for the Qwen3-based predecessor, while hosted Jev still leads at 0.857. He also flags a regression, a five-question request taking 779ms on Qwen3.5 Kev-4B against 174ms on the Qwen3 version, and recommends the older checkpoints until MLX lands. Publishing your own missed pre-registered criteria instead of moving the gate is rarer than it should be.