Kev brings small, local decision models to the Jev-shaped gap

Jared Palmer has released Kev, a family of small AI decision models that return answers such as yes/no probabilities, multiple-choice decisions and ratings instead of conversational prose. The practical point is compellingly unglamorous: developers can run these models locally for fast, repeated decisions without sending every request to a hosted frontier model.

Watch Desk analysis

What happened

Kev is built on Qwen3.5 bases and comes in 0.8B, 4B and 9B versions, according to the project’s GitHub repository. The models are designed to match TypeSafe AI’s System One API and can run on CUDA hardware and Apple Silicon. Palmer’s evaluation data indicates that the newer Qwen3.5-based versions generally improve on the previous Qwen3 generation for accuracy and Brier scores, which measure the quality of probabilistic predictions.

That makes Kev a notable open release rather than merely another chatbot wrapper. Its output is structured for tasks such as classification, routing and scoring, where a number and a confidence estimate may be more useful than a beautifully phrased paragraph. The repository is available on GitHub.

Why it matters

Most software does not need an AI poet. It needs to decide whether an item is spam, which queue should receive a ticket, or whether a search result deserves another look. Small decision models could make those jobs cheaper, faster and easier to keep inside a developer’s own environment.

The local option also changes the trade-off. Teams can experiment without an API bill or a network round trip, while keeping more control over data. The catch is that probabilities are not automatically trustworthy. Palmer’s repository supplies evaluation results, but the evidence does not establish that Kev outperforms larger models across real production workloads.

Our read

Kev is interesting because it narrows the ambition. Rather than pretending every software problem needs a general-purpose conversational model, it targets the fiddly decisions that quietly fill modern systems. That is often where useful infrastructure begins, although “small and local” is not a synonym for “safe and unbiased”.

Developers should test Kev on their own labelled data, compare calibration as well as raw accuracy, and keep a human review path for consequential decisions. The glamorous model launch may get the fireworks, but a dependable classifier is the one that keeps turning up for work.

What to watch

• Independent evaluations:
  whether outside developers reproduce the reported accuracy and calibration results.
• Hardware reach:
  how well the 0.8B, 4B and 9B models run on ordinary laptops and modest GPUs.
• Production use:
  whether teams adopt Kev for routing, ranking and classification beyond toy examples.
• Decision quality:
  whether local speed comes with acceptable performance across languages, domains and edge cases.

Discussion spark: Will small local decision models become a more useful layer of software than general-purpose chatbots, or will their lack of explanations make teams too wary to deploy them?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.