Kev

Overview

Kev is a family of small decision models built on Qwen3.5 and Qwen3.8 and based on the architecture described in Jev's Architecture Unmasked. It ships as Kev 1.0 with four sizes, from a 0.8B model that runs on a laptop to a 27B model for a single data-centre GPU. You can use the pretrained weights or train your own, and the API matches TypeSafe's System One so their Python SDK can point at a local Kev server.

Key Features

• Yes/no, multiple-choice, and rating questions can be asked in one request; questions share the text but cannot read each other.
• Calibrated probabilities are provided by default, with each checkpoint shipping with a fitted temperature.
• The TypeSafe Python SDK works against a Kev server unchanged for drop-in Jev compatibility.
• Four sizes are versioned together as Kev 1.0: Kev-0.8B, Kev-4B, Kev-9B, and Kev-27B.
• Documents of up to 65,536 tokens are supported, with validated context documented per model and execution on CUDA, ROCm, and Apple Silicon through MLX.
• You can fine-tune on your own labelled examples and deploy an HTTPS endpoint that scales to zero when idle.

Use Cases

• Routing support tickets by department and assessing escalation or customer frustration from text.
• Classifying or scoring documents and short text with calibrated probabilities for automation thresholds.
• Fine-tuning decision questions on domain-specific labelled data such as routing categories, escalation rules, or another language.
• Running local or self-hosted decision-model inference for applications that already use the TypeSafe System One API.

Getting Started

• Try the hosted Hugging Face Space to run Kev-4B and Kev-0.8B in the browser with nothing to install.
• For local use, install Python 3.12 or 3.13 and uv, clone the repository, run uv sync --extra serve, then start the server with uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009.
• Send requests to POST /v1/systemone with a state and questions, or use the TypeSafe Python SDK included by uv sync --extra serve.
• Fine-tune with kev.train using JSONL examples, or use the kev-finetune coding-agent skill; deploy with the kev-deploy skill or a Modal app.

Deployment & Requirements

• Local serving requires Python 3.12 or 3.13, uv, and a supported accelerator; the default device selection probes CUDA or ROCm, then MPS, then CPU.
• Kev-0.8B runs on a 4 GB GPU or any Apple Silicon Mac; Kev-4B and Kev-9B target L40S or H100 class GPUs; Kev-27B needs an 80 GB GPU and serves in bf16 at about 66 GB resident with serving buffers.
• For an HTTPS endpoint, a Modal account is required; the deployment scales to zero when idle and the first request after idle waits about 35 seconds for a container to start.
• The local server binds to 127.0.0.1 by default and can require a bearer key with KEV_API_KEY.

Before You Adopt

• License: Apache-2.0. Review its terms before using, modifying, or distributing the project.
• Thresholds should be validated on local data before relying on calibrated confidence.
• Fine-tuning gains are measured in distribution and do not establish performance on unrelated tasks.
• Modal HTTPS deployments scale to zero, but the first request after idle incurs a cold-start wait.
• Kev-27B has substantial serving memory requirements and needs an 80 GB GPU.
• The lower-latency Modal path keeps a paid GPU container warm even when idle.