Build1 publisher2 min readPublished
Ninety-five dollars of rented H100 time produced Kev's three-model Qwen3.5 port
Jared Palmer's Kev now ships as 0.8B, 4B and 9B variants that score typed answer options over frozen Qwen3.5 weights, and on his own development comparison hosted Jev still scores higher than the largest of them.
The Engineer · Build desk
What happened
- Jared Palmer released a Qwen3.5-based generation of Kev on September 20th, making the open-source decision-model project a three-model family developers can train, inspect and run on their own hardware.
- Kev takes a document or other application state, evaluates a set of typed questions, and returns probabilities for yes-or-no, multiple-choice and ordered-score answers.
- Palmer's execution log puts the Qwen3.5 work at roughly $95 of Modal H100 usage plus $0.03 of Jev API calls, covering probes, several training trials, benchmark runs, locked test reads and an ablation.
- On the repository's locked, new-source test set, Kev-9B scored 0.837 accuracy against 0.780 for the prior Kev-8B, with Kev-4B at 0.832 and the 0.8B variant at 0.668.
- The code and model adapters ship under Apache 2.0 alongside frozen evaluation data and model cards, and Kev can be called through TypeSafe's Python SDK with the server and weights local.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- contradiction The repository's own development comparison puts hosted Jev at 0.857 and Kev-9B at 0.812, and Palmer warns about that comparison, so a team choosing self-hosting is doing it on a number the author himself flags.
- constraint Kev never generates explanatory text at inference, so any workflow that needs a written rationale next to the decision has to produce it with a second model or from the probabilities.
- cost The logged spend bought the Qwen3.5 port experiments, and Kev's development predates that port, so budgeting a decision-model family from the $95 figure budgets for a port and not for the family.
- decision Nobody outside the project has reproduced these scores. Whoever adopts Kev carries the evaluation work: the released harness and frozen data are the only way to check the claim against your own tickets.
A call into Kev scores options. Each released checkpoint is a rank-16 LoRA adapter and a small pointer-style readout head attached to a Qwen base model. The head scores the answer options supplied in the request, and a softmax turns those scores into probabilities [14]. None of that inference time goes into explanatory prose [15]. The base weights stay fixed while the adapter and the head learn the decision task [18].
The port cost Palmer his batching trick. Qwen3.5's bases mix conventional attention with recurrent Gated DeltaNet layers, and those layers do not respect the block attention mask the earlier Qwen3 version relied on [16]. Kev now runs each question in an independent row, computing the shared state once and reusing the cache, so questions stay isolated and the input is not re-encoded per question [17].
Training ran two epochs over 12,576 examples: 10,000 from ten public datasets, 896 generated policy examples and 1,680 built from 60 generated rule structures [18][1]. Palmer built Kev with help from Devin, the coding agent made by Cognition, where he is vice president of engineering [5]. His repository says no Jev outputs were used in training [10]. Palmer's execution log itemizes spend down to three cents of API calls, more bookkeeping than most model releases publish [13].
The accuracy table is a claim about Palmer's questions. For it to transfer, your typed questions would have to resemble ten public datasets plus generated policy and rule examples. Your definition of a correct answer would have to match the harness that scored them [18]. Kev-4B sits 0.005 behind the 9B on the locked set [2].
Where Palmer compared the two systems directly, on the repository's development comparison, hosted Jev leads Kev-9B by 0.045 [3], and Palmer explicitly warns about that comparison [24]. The same 9B checkpoint also scores 0.025 higher on the locked, new-source set than it does on the development comparison [4]. The generation-over-generation gain and the gap to hosted Jev come from different questions.
TypeSafe AI introduced Jev and its "System One" category on September 15th, and founder Diogo Almeida described it as a model for fast, typed decisions that software can consume directly, according to runtimewire [7][8]. Kev answers through TypeSafe's Python SDK while the server, weights and training code stay local [9].
The Qwen3.5 family landed five days after the Jev announcement, though Kev's development predates the port [12]. Code and adapters are Apache 2.0, with frozen evaluation data and model cards published alongside [11]. A team that doubts the numbers can rerun the harness that produced them on its own tickets.
What to watch
- An independent reproduction of Kev's locked test-set scores on data the project did not curate.
- Any change to TypeSafe's Python SDK or its terms that affects clients pointing at a self-hosted Kev server.
- A published cost log for Kev's pre-Qwen3.5 development, which would show what the family cost end to end.