Kev

Kev is a family of small decision models built on Qwen architectures that train and run locally for calibrated probability tasks.

Kev is a family of small decision models built on Qwen3.5 and Qwen3.8, based on the architecture described in Jev's Architecture Unmasked. Users can utilize pretrained weights or train their own models on custom labelled examples. The API matches TypeSafe's System One, allowing the Python SDK to work against a local server unchanged. It handles yes/no, multiple-choice, and rating questions in a single request, with each checkpoint shipping with a fitted temperature for calibrated probabilities. Four sizes are available, ranging from 0.8B for laptops to 27B for data-centre GPUs, supporting documents up to 65,536 tokens on CUDA and Apple Silicon through MLX.

Kev at a glance

Best for: Software engineer, Researcher
Runs on: Web, API, CLI, open source
Pricing: Free
Released: September 2026
Category: AI research tool
Website: github.com/jaredpalmer/kev

Key features

Supports yes/no, multiple-choice, and rating questions in one request
Ships with fitted temperature for calibrated probabilities by default
Compatible with TypeSafe's Python SDK for local servers
Available in four sizes from 0.8B to 27B parameters
Processes documents up to 65,536 tokens
Supports fine-tuning on custom labelled examples
Deploys HTTPS endpoints with scaling to zero when idle

Use it for

Routing customer support tickets and evaluating complaint data
Analyzing contracts and processing long documents
Automating decision-making tasks with probability thresholds

Alternatives to Kev

Jev — Fast, cheap typed decisions that software can act on directly
Vev — Runs locally on NVIDIA hardware with a fast, Jev-compatible API.
NanoJev — It runs parallel decisions and outputs probability distributions without output-token decoding.
SemIf OpenJev — Extracts direct token probabilities without generating text or running JSON repair loops.
Jiwo — It provides low-latency decision-making with a 50ms median request time.
eu/jev — Inference hosted in Germany on Hetzner servers with no storage of prompts.
Laya — It runs typed decisions in a single forward pass over 100+ languages.
Clef — Offers low latency and high-speed classification for agentic workflows.
Decido — Runs locally on CPU while beating published benchmark scores.
Bud Decision Studio — Run open decision models locally and get calibrated probabilities in milliseconds.
JEV sees — Turns visual input into program-ready structured choices and probabilities.
Vegaml — Runs locally on your hardware with calibrated probabilities and conformal answer sets.

All Kev alternatives