Developer Jared Palmer released Kev, an open source family of decision models built on Alibaba's Qwen3.5, in three sizes: 0.8 billion, four billion and nine billion parameters, according to its GitHub repository. The project is licensed under Apache 2.0, matching the license of the underlying Qwen models.
Kev is modeled on Jev, the decision-only model TypeSafe AI released last week that answers yes-or-no questions, multiple-choice selections and rating tasks instead of generating text, a launch The Poliety covered at the time. Palmer's README says Kev is "based on the architecture described in Jev's Architecture Unmasked," and benchmarks its results against Jev's published numbers.
Kev answers each question independently: "questions share the input text but can't read each other," the README says, so checking five conditions on the same document produces five separate answers rather than one that reasons across all of them. The project uses LoRA adapters and a pointer-head mechanism to score answer options, and ships a REST API compatible with TypeSafe's own System One format along with a web playground for testing.
Kev is at least the second open recreation of Jev's approach to reach GitHub within days of the original's release, after a Claude Code plugin called fast-jev-compaction adopted the same scoring method for context management. A decision-only model is a narrow enough idea that once TypeSafe published how it worked, replicating it became a weekend project instead of a research problem.