Jared Palmer has released Kev, an open-weight family of small decision models that clone the architecture of TypeSafe's proprietary Jev system. Built as rank-16 LoRA adapters on Qwen3.5 base models, Kev comes in three sizes — 0.8B, 4B, and 9B parameters — and handles yes/no, multiple-choice, and rating questions in a single inference pass. The API mirrors TypeSafe's System One endpoint, meaning their Python SDK works out of the box pointed at a local server. The core mechanic is structured: you send a block of text (a support ticket, a document, a chess position) and a set of typed questions. The model returns calibrated probabilities rather than single labels. A customer complaint mentioning a return, a late delivery, and a billing issue comes back with probability distributions across all three departments. Each question shares the input context but is masked from the others via a custom attention scheme, preventing cross-contamination between answers. Performance lands close to the closed commercial system. Kev-9B scores 0.852 accuracy on held-out test data from unseen sources, trailing Jev's 0.857 on a development set — though the comparison is imperfect since Jev's training data is unknown and it hasn't been evaluated on the same test set. The 4B model hits 0.837, and even the 0.8B model manages 0.684 on new sources. Brier scores (lower is better) range from 0.237 for the 9B to 0.460 for the 0.8B, indicating the larger models produce well-calibrated confidence estimates. Calibration is handled by a learned temperature parameter (roughly 2.1–2.4) stored with each checkpoint. This post-hoc scaling drops Kev-9B's calibration error from 0.106 to 0.042 and cuts high-confidence errors from 8.7% to 4.0%, approaching Jev's 3.7%. The accuracy numbers remain identical whether you use the calibrated or raw logits. A notable pragmatic feature: setting KEVDATEFACTS=1 pre-computes date arithmetic and injects it into the prompt, boosting deadline-policy accuracy from 0.80 to 0.90. The release is unusually complete for an open-source ML project. Training code, evaluation data, model cards with confidence intervals, and a web playground are all included. The playground lets users test option-order sensitivity, compare packed versus separate inference, and probe question isolation — the kind of diagnostic tooling that matters for trust in production. There's even a chess demo where the board state is the input and legal moves are choice options. The controlled comparison between Qwen3 and Qwen3.5 generations is the most informative detail for the ML community. Same data, same training settings, different base: Kev-9B gained 7.3 accuracy points over its Qwen3 predecessor (95% CI +2.8 to +11.7). The 4B gained 2.9 points and the 0.8B gained 4.8 points. This isolates the base model's contribution and gives the field a rare clean signal about what newer foundations buy you in fine-tuned downstream tasks. The strategic picture is straightforward: this is commoditization of a proprietary API. TypeSafe charges for Jev access; Kev lets you run a near-equivalent locally on a 32GB Mac in bf16 at 495ms latency. Whether this pressures TypeSafe's pricing, accelerates adoption of decision models as a category, or simply provides an escape hatch from vendor lock-in depends on how production teams weigh the 3.5-point accuracy gap against the cost and control benefits of self-hosting.
