Jared Palmer describes his new project in about as few words as a project can be described: a tiny, Jev-like family of decision models built on top of Qwen3.5 that you can train and run on your own. That’s the whole pitch. No launch video, no waitlist, no pricing page. Just a repository, a claim, and an invitation to try it yourself.
My first reaction was relief. So much AI news arrives pre-chewed, wrapped in language designed to make you feel like you’re already behind. “Train and run on your own” is the opposite of that. It’s a sentence that assumes you have a computer and some curiosity.
What a decision model actually does
If you’ve only used chatbots, the phrase “decision model” might not land. So here’s the distinction, borrowed from how people have been explaining Jev, the project Kev takes after. As Tonbi’s AI Garage put it in a walkthrough of Jev, the model gives up free-form text.
That’s the trade. A normal language model answers you with prose. It explains, hedges, offers three options, apologizes, and circles back. A decision model drops most of that and produces a choice instead. Think of it less like a writer and more like a switch operator: something comes in, a call gets made, the thing moves down the right track.
Why would anyone want less text? Because if you’re building an AI agent, most of what happens under the hood isn’t conversation. It’s hundreds of small judgment calls. Is this email spam or a real customer? Does this support ticket go to billing or engineering? Should the agent search the web or answer from what it already knows? Each of those is a decision, and paying for a chatty model to produce five paragraphs before picking an answer is slow and expensive.
The size thing matters more than it sounds
Kev comes in three sizes: 0.8B, 4B, and 9B parameters. Parameters are roughly the number of internal dials a model has, and the count is the closest thing we have to a shorthand for how big it is.
For scale, the models behind the well-known commercial chat assistants are widely believed to be hundreds of times larger than 9B. Those need data centers. A 0.8B model is small enough to be an afterthought on modern hardware.
Small brings real advantages for the kind of work described above:
- It runs on your machine. Your data doesn’t leave the building, which matters if you’re handling anything private.
- It’s fast. A decision that takes a fraction of a second is a decision you can make thousands of times without waiting.
- It’s cheap. No per-request billing on hardware you already own.
- You can train it. This is the part I find most interesting, and I’ll come back to it.
Built on top of Qwen3.5
Kev isn’t built from scratch. It sits on top of Qwen3.5, an existing open model family, and Palmer’s own description notes it’s similar to Jev but uses different training methods.
The useful thing to understand here is that modern AI work is mostly remixing. Someone spends enormous resources training a base model that has general language ability. Other people then take that base and shape it toward a specific job. Kev is the second kind of work: take a capable open model, train it differently, aim it at decisions.
Which is also why “different training methods” is doing a lot of quiet work in that description. Two projects can start from the same base and land in very different places depending on how they’re taught.
A word about the clone wave
Kev landed on Hacker News and did well, over 370 points and more than 160 comments within a day. But there’s context worth carrying into this.
The AI site explainx.ai published a fairly candid admission about its own coverage: it had written up six Jev clones that shipped in 48 hours, and that coverage was itself assembled from digest headlines with no primary source to check against. Its word for the resulting confusion was that it “wasn’t unreasonable.”
I appreciate the honesty, and I want to be straight with you about the same limit. A repository description and a hot comment thread tell you a project exists and that developers are curious. They don’t tell you it works well. Benchmarks, independent testing, and people actually shipping things with it are what turn an interesting repo into a tool you’d rely on.
Why I’d still keep an eye on this
Set aside whether Kev specifically becomes important. The shape of it points somewhere.
For most of the past few years, “using AI” meant renting access to something enormous that lived somewhere else. Kev belongs to a different pattern: small, specialized, local, and trainable by whoever downloads it. The person building an agent to sort their company’s invoices doesn’t need a model that can write poetry. They need something that gets one call right, quickly, on their own hardware.
Whether you ever touch it or not, that direction is the more interesting story. Smaller models doing narrower jobs, on machines their owners control.
🕒 Published: