Kev

Kev is a family of small open-weight decision models that answer several typed questions about a piece of text or application state in one request. Jared Palmer published it as a personal open-source project, with models at 0.8B, 4B and 9B parameters built on Qwen3.5.

What Kev does

• Yes or no questions: probabilities for yes and no answers.
• Choices: a probability for each option in a multiple-choice question. Option order can change answers.
• Ordered ratings: scores on an ordered scale.
• Drop-in server: implements the same System One interface as TypeSafe's hosted Jev service, so TypeSafe's Python SDK can point at a local Kev server.

Customers and markets

Developers who want a small local model for typed answers. The author advises testing on their own data before choosing thresholds.

Key facts

• Licence: Apache 2.0 for the code, adapters and evaluation data.
• Hardware: CUDA, ROCm and Apple Silicon are supported. On a 32 GB Mac the 4B and 9B models run in bf16, slowly, and the server handles one request at a time.
• Author's accuracy: Kev-9B scores 0.852 on a held-out test set, measured by the author's own harness.
• Context: the training examples were at most 1,024 tokens per state and question, and the server accepts up to 8,192.
• Release: published as a GitHub release on 20 September 2026, with the weights updated on 21 September.