Kev: Tiny Jev-like family of decision models built on top of Qwen3.5

2026-09-21 | original

read original get Apple Mac Studio (M2 Max) → more articles

Why This Matters

Kev is an open-source, small-scale decision model family built on Qwen3.5 that lets developers run and train lightweight classification/routing models locally instead of relying on large closed APIs. This matters because it lowers the barrier for building fast, cheap, self-hosted decision logic (like support ticket triage) while remaining compatible with existing commercial SDKs, giving developers more control and reducing vendor lock-in.

Key Takeaways

Kev offers 0.8B, 4B, and 9B parameter models with open training code and evaluation data, runnable on both CUDA and Apple Silicon.
It supports mixed question types (yes/no, multiple-choice, rating) in a single request and is API-compatible with TypeSafe's System One SDK.
The project emphasizes accessibility and transparency, letting developers self-host and fine-tune decision models rather than depending solely on proprietary services.

Worth a Look

Apple Mac Studio (M2 Max) — Since Kev's 4B and 9B models run natively on Apple Silicon and fit comfortably in 32GB using bf16, a Mac Studio is a great local box for experimenting with these decision models without needing a discrete GPU. It's a solid choice for developers who want to train, serve, and tinker with small LLMs like Kev entirely on-device.

See Apple Mac Studio (M2 Max) on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.

Kev

Small Jev-like decision models you can train and run yourself.

Kev is a family of small decision models built on Qwen3.5 and based on the architecture described in Jev's Architecture Unmasked. You can use the pretrained weights or train your own. The API matches TypeSafe's System One, so you can point their Python SDK at your local server.

Highlights

0.8B, 4B, and 9B models, with training code and evaluation data.

Yes/no ( noul ), multiple-choice ( choice ), and rating ( score ) questions in the same request.

), multiple-choice ( ), and rating ( ) questions in the same request. Questions share the input text but can't read each other.

Runs on CUDA and Apple Silicon. The 4B and 9B models fit a 32 GB Mac using bf16; see Serving Performance for what to expect on a Mac.

A web playground for trying your own inputs and checking how option order affects the answers.

Quick Start

... continue reading

Explore topics: kev qwen3.5 decision models typesafe system one apple silicon