Remember the rhythm of the last few years? A lab ships something that behaves strangely well, nobody outside can run it, and within a couple of weeks a handful of small reproductions appear on GitHub with names that sound like someone’s cousin. The reproductions are almost never as good. They are almost always more useful, because you can open them.
Kev is the newest entry in that tradition. It’s a small family of decision models built on top of Qwen3.5, positioned as Jev-like but trained differently, shipped at 0.8B, 4B, and 9B parameters, and explicitly meant to be trained and run on your own hardware. It landed on Hacker News at 370 points and 164 comments, which in 2026 is roughly the signature of “people want this to be real.”
What Kev is, and what it isn’t
The framing matters here. Kev is not pitched as a general assistant. It’s a decision model — the same category Jev popularized by trading free-form text generation for structured decision output. That’s a meaningful architectural commitment, not a packaging choice. A model whose output space is a decision rather than a paragraph has a different loss surface, different failure modes, and a very different integration story for anyone building agents.
For those of us who spend our time on agent architecture, this is the part worth studying. Most agent stacks today paper over a mismatch: they use a text generator as a controller, then spend enormous engineering effort parsing, validating, and retrying its output. Every schema-enforcement library and every JSON repair function in your dependency tree exists because of that mismatch. A model trained to emit decisions directly collapses a layer of that stack. Whether it collapses it well is the empirical question.
The size ladder is the real design statement
0.8B, 4B, 9B. Those three numbers tell you the intended deployment story more clearly than any README could.
- 0.8B is edge and in-loop territory. Something you can call thousands of times inside a single agent trajectory without thinking about cost or latency budgets.
- 4B is the consumer-GPU sweet spot — fine-tunable on hardware a single researcher owns, which is what makes iteration on training method possible at all.
- 9B is the ceiling where you still fit comfortably on one accelerator but get enough capacity for harder routing and tool-selection decisions.
Read that ladder as a hypothesis: decision-making, as a task, may not need the parameter count we’ve been throwing at it. If a controller only has to pick among bounded options given structured context, the capability curve might flatten early. That would be a genuinely important result for agent design, because it would mean the expensive model belongs at the leaves of your system, not at the root.
Same base, different training — that’s the variable to watch
Kev and Jev reportedly share a family resemblance but diverge in training method. Since Kev builds on Qwen3.5 as its substrate, the base model is largely a controlled variable. What varies is how the decision behavior gets installed.
That’s an unusually clean experimental setup for open work, and it’s the thing I’d want reproduced first. Decision models live or die on calibration, not just accuracy. A controller that is confidently wrong 5% of the time is worse for an agent loop than one that is uncertain 20% of the time and says so, because the uncertain one can escalate. Training method is exactly where that property is decided.
The repo also carries planning documents — PLAN.md, PLAN_Qwen35.md — with in-progress notes including an MMLU-Pro figure attached to a 35B correction. I’d treat those as lab notebook entries rather than published results, and I’d encourage everyone else to do the same. Working notes are a gift to readers precisely because they’re unpolished. They stop being a gift the moment someone quotes them as a benchmark.
A note on how this story got reported
There’s a detail in the coverage I find more instructive than the model itself. explainx.ai, writing about Kev, acknowledged that its own earlier piece on six Jev clones shipping in 48 hours had been assembled from digest headlines with no primary source to check against.
That’s an honest admission, and it describes the actual information failure mode of this moment. Small models now ship faster than anyone can evaluate them, so coverage gets built from other coverage. The citation graph turns into a loop with no ground truth in it. Six clones in two days is not six validated approaches — it’s six repositories, some of which may be substantially the same idea.
What I’d actually measure
If you want to know whether Kev is interesting rather than merely trending, clone it and run three things. Decision accuracy across the three sizes on your own task distribution, to find where the curve flattens. Calibration under distribution shift, because that’s what breaks in production agents. And training cost to recover the behavior on the 4B, since that number determines whether the “train it yourself” claim is real or aspirational.
The fact that you can run those experiments is the point. A local, small, openly trainable decision model is a research instrument, not just a deployment option. That’s a better reason to pay attention than the vote count.
🕒 Published: