Skip to content

Released Sep 2026

Record sheet · No. 145

Kev

A family of small models that read one document and answer a typed set of questions with a probability for every option. No text is generated.

GitHub avatar of jaredpalmer, not the project’s own logo — taken from github.com on 2026-10-02.

Screenshot of Kev
Editor screenshot, 2 Oct 2026Kev ↗

What it is

An application usually asks the same handful of questions over and over: which queue is this, is it urgent, does it need a person.

Kev turns each one into a typed question with a fixed set of options and returns a probability for every option in a single forward pass, so the caller gets a number it can threshold rather than a paragraph it has to parse. Four sizes ship together as Kev 1.0, all on Qwen bases: a 0.8B that runs on a laptop, then 4B, 9B and 27B for progressively larger GPUs.

The API matches TypeSafe’s System One, so their Python SDK points at a local Kev server unchanged.

Who built it

The repository is effectively one person’s: 312 of its 333 commits belong to Jared Palmer, eight other accounts contribute one each, and a Devin integration bot accounts for twelve. The README describes the project as built with Devin.

How it is put together

The parts · 6

One base model, one adapter, one forward pass. A request carries a document, which Kev calls the state, plus a set of typed questions, each with a fixed list of options. The model returns a probability for every option of every question, and a pointer head reads each answer out of a single <decide> token, so nothing is generated word by word. Questions share the state but cannot read each other, and the Space demonstrates that a secret written into one question is invisible to its siblings. Each checkpoint ships with a fitted temperature: on Kev-4B it takes out-of-domain calibration error from 0.12 to 0.05 without changing the top answer.

The state
One document per request, up to 65,536 tokens. Every question sees it. No question sees another question.
Typed questions
Three kinds: choice with named options, noul for a yes or no, and score with an ordered list of levels. Each one comes back as a probability per option.
The readout
A pointer head scores each question’s <decide> token against the spans of its options. The answer is read out of the model, not written by it.
The adapter
LoRA r=16 over the attention, MLP and Gated DeltaNet projections of a frozen Qwen base, merged into the base weights in fp32. Kev-27B is fine-tuned in full instead.
The API contract
Requests and responses follow TypeSafe’s System One, so their Python SDK works against a local Kev server without changes.
The evaluation suites
Frozen per round and published as jaredpalmer/kev-suites, with a template that puts every training run’s own questions into the same form.

Choices, and the alternative

  • Answer with a probability distribution over generating text and parsing it

    Stated as the design, not as a preference: one document and typed questions in, a probability for every option out, one forward pass, no text generated.

  • Release four sizes under one version over shipping them separately

    The four models are released together as Kev 1.0 and every Hub repository carries the v1.0 tag, so a pinned tag always loads the same weights.

  • One training recipe across the three smaller sizes over a recipe per size

    The README states that the 0.8B, 4B and 9B share one recipe, a small adapter on a frozen base. The 27B is the documented exception, fine-tuned in full.

Read fromREADME, space/README.md, the model table and the repository tree of jaredpalmer/kev, read 2026-10-03.

Build log

3 stages
  1. 01

    Four sizes released at once as Kev 1.0

    The repository was created on 2026-09-17 and four models shipped together: 0.8B, 4B, 9B and 27B. Each Hub repository carries a v1.0 tag, so a pinned tag always loads the same weights, and the three smaller checkpoints are attached to the GitHub release with SHA-256 checksums. The 27B is the exception to the family recipe. The 0.8B, 4B and 9B are a small adapter trained on a frozen Qwen base; the 27B starts from Qwen’s post-trained release and is fully fine-tuned, which is why it ships as 51 GB of weights instead of an adapter.

  2. 02

    Every evaluation round is kept

    The tree runs to 3,049 files, and most of its weight is results rather than code: evals/round5 at 14.6 MB and evals/devtools-v1 at 12 MB, with runs/r5-combined behind them. Each round keeps its own data and measurements rather than overwriting the last one, so a change to the recipe can be compared against the round before it after the fact. The research notes sit next to the artefacts they produced: docs/autoresearch.md, 22 KB. The suites are published separately as jaredpalmer/kev-suites, which is what lets a result be re-read against the data it came from.

  3. 03

    How much of it Devin wrote

    The repository holds 333 commits and Jared Palmer owns 312 of them. 145 commits carry a co-author trailer, and 134 of those name Devin: 127 as Devin, six as Devin AI, one as the integration bot. The remaining eleven name a person or a Claude model. The README says the same thing in five words, “Built with Devin”. What the trailers support is that Devin worked inside the development loop for most of the history. Which changes it made by itself is not something a trailer can answer, and this record does not claim to know.

What they would tell you

  • Kev-0.8B is a sub-1B model and noticeably weaker out of domain. The Space says so without being asked, and adds the instruction that matters: measure on your own inputs.
  • Calibrated is not the same as correct. Kev-4B sits at 0.12 out-of-domain calibration error raw and 0.05 with the fitted temperature; the top answer does not move either way.
  • A model trained on the questions you have is not a model that answers the questions you have not thought of yet. Product-shaped questions with no training analogue are not guaranteed.
  • The README compares Kev-27B with the hosted Jev and then says what the comparison is worth: within a point of it on new sources, but nobody outside TypeSafe knows what Jev was trained on, so this is not a controlled test of two architectures.

Revision log

1

This record is kept current rather than frozen: when the project moved and a paragraph stopped being true, the paragraph was corrected in place. Every corrected paragraph carries a marker, and each one appears here with what it said before.

  1. 2026-10-03RecordRewritten for a reader who has not heard of the project. The mechanism now comes first and the commit history last, the headline no longer carries a commit count, and the sentence claiming one person directing an autonomous agent has been replaced with what the trailers actually show. The tooling was corrected from Claude to Devin, which is what 134 of the 145 co-author trailers name.

Adjacent records

All records →