● Covered by 2 sources · 2 reportsLow impact2 neutral
Kev: New Family of Small Decision Models Based on Qwen3.5 Architecture Released
Key points
- Kev models are 0.8B, 4B, and 9B parameters.
- Built on Qwen3.5, based on Jev's Architecture Unmasked.
- Supports yes/no, multiple-choice, and rating questions.
- Runs on CUDA and Apple Silicon; API matches TypeSafe's System One.
- Jared Palmer released Kev.
- Kev is prefill-only.
- Kev processes state, questions, and candidates in a single forward pass.
- Kev reads decisions from a pointer head without an autoregressive decoding loop.
Introduction to Kev Decision Models
Kev is a newly released family of small decision models. These models are built upon the Qwen3.5 architecture and draw inspiration from the design principles outlined in Jev's Architecture Unmasked. Users can either utilize the provided pretrained weights or train their own custom models.
Model Sizes and Capabilities
The Kev family includes models with 0.8 billion, 4 billion, and 9 billion parameters. They are designed to handle various question types, including yes/no (noul), multiple-choice (choice), and rating (score) questions within a single request. Questions share the input text but operate independently.
Technical Specifications and Compatibility
Kev models are compatible with both CUDA-enabled systems and Apple Silicon. The 4B and 9B models can run on a 32 GB Mac using bf16 precision. The API for Kev models matches TypeSafe's System One, allowing their Python SDK to be directed at a local Kev server. A web playground is also available for testing inputs and observing the impact of option order on answers.
Local Deployment and Usage
To run Kev locally, users need Python 3.12+ and uv. The process involves cloning the GitHub repository, installing dependencies, and starting the server with a command like `KEV_DTYPE=bf16 uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009`. This command downloads the necessary adapter and base model on the first run. The `--run` argument also accepts local checkpoint directories or Hub revisions.
Example Interaction
An example demonstrates sending a customer service ticket to a locally running Kev-4B model. The model processes the input and provides answers for department assignment, escalation necessity, and customer frustration level, including confidence scores and probabilities for each response.
Updates
- Jared Palmer released Kev.
- Kev is prefill-only.
- Kev processes state, questions, and candidates in a single forward pass.
- Kev reads decisions from a pointer head without an autoregressive decoding loop.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Primary sources
GitHub cmc_internal/api GitHub github/collect GitHub jaredpalmer/kev GitHub _private/browser GitHub get-started/accessibility GitHub open-source/sponsorsHow outlets covered it
Developer Jared Palmer released Kev, a new family of open decision models built on Qwen 3.5, which aims to reduce token consumption in AI agents by making decisions without generating text. This approach addresses the issue of AI agents spending significant compute on unnecessary text generation for internal decisions, potentially lowering operational costs for AI workloads.
Kev is a new family of small decision models, available in 0.8B, 4B, and 9B sizes, built on Qwen3.5 and inspired by Jev's Architecture Unmasked. These models support yes/no, multiple-choice, and rating questions, and can be run locally on CUDA and Apple Silicon, with an API compatible with TypeSafe's System One.