Jared Palmer ports Kev to Qwen3.5 for roughly $95 in H100 time

Why it matters

Kev turns a newly commercialized model category into open infrastructure developers can run locally. Palmer's roughly $95 Qwen3.5 port also shows how tightly scoped models, rented GPUs and coding agents can compress model development without hiding the benchmark gaps.

Jared Palmer (@jaredpalmer) released a Qwen3.5-based generation of Kev on September 20th, turning the open-source decision-model project into a three-model family that developers can train, inspect and run on their own hardware.

The release includes 0.8B, 4B and 9B parameter variants. 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. It is designed for the small semantic judgments buried inside software workflows: routing a support ticket, choosing an escalation path or scoring a customer's frustration.

Palmer built Kev with assistance from Devin, the coding agent made by Cognition, where he is vice president of engineering. His path to Kev runs through a string of developer products: he founded Turborepo, which Vercel acquired in December 2021, and his biography credits him with creating Vercel's v0 interface generator and AI SDK. He later held senior roles at GitHub and Xbox before joining Cognition.

Kev fits that history. Palmer has repeatedly taken infrastructure that sits behind a large product and repackaged it as something developers can run, modify and build around. This time, the target is a new model category that had barely been public for a week.

A model that answers instead of writes

TypeSafe AI introduced Jev and its "System One" model category on September 15th. Founder Diogo Almeida described Jev as a model for fast, typed decisions that software can consume directly. Instead of generating a paragraph and asking an application to parse it, Jev returns values and probabilities from a schema supplied in advance.

Kev follows that interface and can be called through TypeSafe's Python SDK, while keeping the server, weights and training code local. Palmer's repository says no Jev outputs were used in training. The code and model adapters are available under Apache 2.0, along with frozen evaluation data and model cards.

The Qwen3.5 family arrived five days after TypeSafe's public Jev announcement, although Kev's development predates the port. Palmer's execution log records roughly $95 of Modal H100 usage for the Qwen3.5 experiments, plus $0.03 in Jev API calls. That bill covered probes, multiple training trials, benchmark runs, locked test reads and an ablation. It is a useful counterweight to the capital-heavy framing around model development: Kev narrows the job enough that one experienced builder, an agent and rented GPUs can produce a usable open implementation.

Each Kev checkpoint consists of a rank-16 LoRA adapter and a small pointer-style readout head attached to a Qwen base model. The model scores the supplied answer options directly and applies a softmax to produce probabilities. Kev does not spend inference time generating explanatory prose.

Qwen3.5 forced Palmer to change how the questions are processed. The base models combine conventional attention with recurrent Gated DeltaNet layers, which do not respect the block attention mask used in the earlier Qwen3 version. Kev now runs each question in an independent row, computes the shared state once and reuses its cache. That preserves isolation between questions while avoiding repeated work on the input.

The released models were trained for two epochs on 10,000 examples from ten public datasets, 896 generated policy examples and 1,680 examples built from 60 generated rule structures. The base weights remain fixed while the adapter and pointer head learn the decision task.

The benchmark gap remains

Palmer's published numbers show meaningful gains over Kev's prior Qwen3 generation, with caveats that make the results easier to trust.

On the repository's locked, new-source test set, Kev-9B scored 0.837 accuracy, compared with 0.780 for the prior Kev-8B. Kev-4B reached 0.832, up from 0.806 for its Qwen3 predecessor. The 0.8B model reached 0.668, against 0.642 for the earlier 0.6B checkpoint. These figures come from Kev's own evaluation harness and have not been independently reproduced.

Hosted Jev still scored higher on the repository's development comparison, at 0.857 accuracy against 0.812 for Kev-9B. Palmer explicitly warns that this is not a controlled architecture comparison because Jev's training data is unknown. Kev's original pre-registered development criteria also were not met, even though the locked test favored the new checkpoints. The repository publishes both outcomes rather than quietly replacing the gate after seeing the data.

The release also carries a practical regression. On an Apple M5, the repository measured a five-question request at 779 milliseconds for the Qwen3.5 Kev-4B, compared with 174 milliseconds for its Qwen3 predecessor. Kev-9B took about two seconds, versus roughly 300 milliseconds for Kev-8B. Palmer recommends the older Qwen3 models for low-latency Mac deployments until an MLX backend is ready.

Those limitations define Kev's immediate use. The project gives engineers an inspectable implementation for experimenting with typed probabilistic decisions, local data and domain-specific fine-tuning. It is not yet a drop-in performance replacement for Jev, and the repository's own results keep Jev ahead on its primary transfer suite.

Palmer keeps the work in public

Kev's release process is unusually legible. The repository contains the training recipes, dataset manifests, experiment plans, confidence intervals and failed decision gates alongside the successful checkpoints. Palmer recorded where the Qwen3.5 base helped, where training erased a date-arithmetic advantage and where the models became confidently wrong on questions that could not be answered from the supplied evidence.

That last problem is already on the work list. Palmer plans to add training examples with uniform targets for unknowable questions, improve date handling, cover assertion-style instructions and build faster Mac serving. The repository had attracted about 1,300 stars by September 21st, suggesting developers are interested in the model category even while its commercial demand remains unproven.

Devin's role also extends Cognition's pattern of using its coding agent on concrete engineering projects and publishing the artifacts. RuntimeWire previously reported that Devin migrated Cognition's website and recorded its own verification runs. Kev goes further by leaving the resulting model code, checkpoints, evaluation harness and experiment log available for outsiders to inspect.

Palmer has shipped enough developer infrastructure to understand that distribution often starts with an artifact people can run. Kev gives the newly visible decision-model category exactly that: a local server, open weights and enough disclosed failure data for engineers to decide where the model belongs in real software.