kev
Family of small Jev-like decision models you can train and run yourself, with a server compatible with TypeSafe's SDK.
Overview
Kev is a family of small decision models built on Qwen3.5 and Qwen3.8, following the architecture described in the 'Jev's Architecture Unmasked' write-up. It answers noul, choice and score questions in one request, where questions share the text but cannot read each other, and returns calibrated probabilities using a fitted temperature shipped with each checkpoint. Its server speaks TypeSafe's System One API, so the TypeSafe Python SDK works against it unchanged.
Kev 1.0 comes in four sizes: 0.8B (runs on any Apple Silicon Mac or a 4 GB GPU), 4B (recommended starting point), 9B and 27B for an 80 GB GPU, with documents up to 65,536 tokens on CUDA or MLX and model cards stating where accuracy drops. Held-out index scores rise from 23.3 for 0.8B to 41.0 for 9B. You can fine-tune on your own labelled examples, with a coding-agent skill that runs the loop on Modal, and deploy a scale-to-zero HTTPS endpoint with one command. A browser demo is on Hugging Face, and evaluation suites are published. Earlier list entries describe smaller Qwen2.5-based versions.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Any yes/no question about the text | — |
| Choice | Which option fits the text? | — |
| Score | Where does the text fall on the rating scale? | — |
How it works
- Download a Kev checkpoint or train your own
- Run the local server (CUDA or MLX) or deploy an endpoint
- Send state and typed questions with the TypeSafe SDK
- Use calibrated probabilities in your code
- Fine-tune on labelled data if needed
Strengths
- Self-hosted, drop-in for Jev's API
- Runs on laptops at the small end
- Long-context support with documented limits
- Training and deployment tooling included
Caveats
- Smaller models trail larger ones on accuracy
- 27B needs data-centre GPUs
- Accuracy versus hosted Jev depends on task
Getting started
Try the Hugging Face Space at huggingface.co/spaces/jaredpalmer/kev, or download weights from the jaredpalmer/kev collection and run the local server.At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- CUDA or MLX server
- Has code
- yes
- Has tests
- yes
- License
- Apache-2.0
- Stars
- 9,079
Use cases and integrations
How curated lists describe it
Trainable replica: a family of small Jev-like decision models on Qwen2.5 at 0.6B, 4B, and 8B that train and run on a MacBook, shipped with their own research runs and evaluation scripts.
yibie/awesome-jevQwen3.5 decision models (0.8B, 4B, 9B) you can train and serve yourself. Choice, Score, and Noul in one forward pass, with published weights and frozen eval suites, and a local server that speaks `/v1/systemone`.
AnotiaWang/awesome-decision-modelsTiny Jev-like decision head built atop Qwen2.5-0.5B that can be trained and run locally on Apple Silicon MacBooks.
logicrw/awesome-jev-projectsJev-like family of decision models built on top of Qwen3.5/3.8 you can train and run on your own
wh000wh000/awesome-jev-liveQwen2.5-0.5B adapter and decision head with training code, released weights, and parallel typed-question inference.
cobanov/awesome-jevJev-style decision models from 0.5B to 8B, built as LoRA adapters on Qwen and served behind a Jev-compatible `/v1/systemone` API.
MrJev/awesome-jev
Frequently asked questions
Can I use the TypeSafe SDK with Kev?
Yes, the server matches the System One API so the SDK works unchanged.
Which size should I start with?
The README recommends Kev-4B, with 0.8B when size matters more than accuracy.
Can I fine-tune it?
Yes, on your own labelled examples, with a skill that runs training on Modal.
Related projects
- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython 11 MB CPU-only model that answers choice, noul and score questions with fitted calibration and no LLM at inference.
ProjectresearchPythonApache-2.0 local decision model (0.8B Qwen fine-tune on llama.cpp) behind the Jev request shape, calibrated and CPU-only.
ServicebetaPythonOpen 596M and 4B decision models that answer Choice, Score and Noul in one pass locally, with calibrated confidence for escalation.
ServicebetaPythonApache-2.0 118M multilingual decision model for Choice, Score and Noul on the System One API, with conformal abstain.
ModelbetaPythonOpen Apache-2.0 System One decision models (4B, 9B, 27B on Qwen) with trainer, data builders, evals and every run record.
ServiceresearchPython