User repository on GitHub
Jev-like family of decision models built on top of Qwen3.5/3.8 you can train and run on your own
- Forks
- 625
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Practices & Patterns
Kev — jaredpalmer
Apache-licensed, locally runnable Jev-style decision models (0.8B, 4B, 9B on Qwen3.5) with released weights, training code, a System One-compatible server, frozen eval suites, and a playground.
- Stars
- 9.3k
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 3 min
- Published by
- Community
- Author
- jaredpalmer
- Last verified
- Sep 2026
About this project
Small Jev-like decision models you can train and run yourself. 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. You can use the pretrained weights or train your own. The API matches TypeSafe's System One, so you can point their Python SDK at your local server. Yes/no (noul), multiple-choice (choice) and rating (score) questions in one request. The questions share the text but can't read each other.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
Highlights
- Weights for Kev-0.8B, Kev-4B, and Kev-9B are released on Hugging Face with training code and evaluation data.
- Its API matches TypeSafe's System One, so the official Python SDK can point at a local server; requests mix noul, choice, and score questions.
- Runs on CUDA, ROCm, and Apple Silicon; the 4B and 9B fit a 32 GB Mac in bf16, and a web playground permutes option order.
- On its frozen new-source test set the author reports Kev-9B at 0.852 versus 0.857 for hosted Jev, but Jev's training data is unknown.
- Fine-tuning gaps are documented: the first Kev-9B dropped deadline questions to 0.72 and KEV_DATE_FACTS=1 recovers 0.90.
Quickstart
bash
git clone https://github.com/jaredpalmer/kev.git && cd kev
uv sync --extra serve
KEV_DTYPE=bf16 uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009Watch out
Apache-2.0, with the Qwen bases under the same license. Not a controlled architecture comparison because Jev's training data is unknown; needs Python 3.12+ and uv, and calibration and option order should be tested on your own data.
Questions about this entry
Short answers drawn from the linked project or page.
Yes. Kev matches TypeSafe's System One request and response shapes, and the TypeSafe Python SDK ships with the serve extra, so you repoint base_url at the local server. The supported question types are noul, choice and score.
On new-source test sets the author reports Kev-27B at 0.851 against Jev's 0.857, with the 4B and 9B within four points. Jev's training data is unknown, so this is not a controlled architecture comparison, and Jev was only run on the development sets of those two suites.
A laptop is enough for the smaller sizes: 0.8B on any 4 GB GPU or any Apple Silicon Mac, 4B on a 32 GB Mac in bf16. Mac support for the 9B and the 27B is listed as expected, not measured.
Validated context is 8,192 tokens for the 0.8B, 4B and 9B and 65,536 for the 27B. That figure is where accuracy on real contracts stays within 3 points of the same model's 8k score at the 95% lower bound; each model card breaks the measurement down by length.
Clone the repo and run uv sync --extra serve, then uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009. It needs Python 3.12 or 3.13 and uv. Note that pip install kev installs an unrelated package, because this project is not on PyPI.
The README says start with Kev-4B, move to Kev-9B if you have a bigger GPU, or to Kev-27B if you have an 80 GB GPU, and use Kev-0.8B when size matters more than accuracy. Weights are on Hugging Face, released together as Kev 1.0 with a v1.0 tag per repo.
Yes. You can fine-tune on your own labelled examples, and a coding-agent skill runs the whole loop on Modal from finding your questions to serving the result. A one-command deploy gives an HTTPS endpoint that scales to zero when idle.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Kev: a tiny Jev-like family built on Qwen3.5
tosh458 points · 199 commentsKev: Tiny Jev-like family of decision models built on top of Qwen3.5Read the thread on Hacker NewsHacker News: Opper hosts Kev 4B and typesafe/jev-latest on one systemone compat route.
felix08912 pointsJev vs. Kev: open-source Jev alternative tested side by sideRead the thread on Hacker NewsX: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Other listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
Self-hosted Apache-2.0 decision engine with a Jev-style wire format that turns a model you point at it into typed choice, score, and noul answers.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
arjun988/KevTypeScript
arjun988/KevTypeScriptMore Jev resources like this
An independent open reproduction of the System One model class: Qwen3.5-based 2B and 35B mixture-of-experts models that return typed Choice, Score, and Noul probabilities in one forward pass, with nothing distilled from Jev.
Jev Practices & Patterns#community#python#open-models
Community
Mapika/deciderPython
Mapika/deciderPythonA from-scratch one-pass option scorer that maps a context and a changing list of text options to one probability per option, adding a hashed n-gram encoder, rival-aware attention, a gated head, and temperature scaling; JSONL data is drop-in compatible with jevlike.
Jev Practices & Patterns#community#python#calibration
Community
olanotolu/jevbetterPython
olanotolu/jevbetterPythonAn independent baseline that reads typed option probabilities straight from a frozen Qwen3.5-4B's logits in one forward pass, reproducing Jev's interface pattern with open models rather than Jev's undisclosed model or training.
Jev Practices & Patterns#community#python#open-models
Community
TheoLeeCJ/SemIfPython
TheoLeeCJ/SemIfPythonA ModernBERT-based open decision model that evaluates Choice, Score, and Noul schemas in one non-autoregressive pass; Verdict 2.0 reports 77.10% accuracy on 2,000 held-out enterprise decisions.
Jev Practices & Patterns#community#python#open-models
A self-hosted implementation of TypeSafe's Jev System One API powered by the 400M-parameter GLiFormer encoder: it serves choice, score, and noul and drops into the official typesafe-sdk via TYPESAFE_BASE_URL, but trails Jev on reasoning-heavy tasks.
Jev Practices & Patterns#community#python#open-models
Community
logan-markewich/jeffPython
logan-markewich/jeffPythonAn open re-creation of Jev's decision model on an open-weights Qwen3.5-4B: it answers a list of typed questions in one forward pass over POST /v1/systemone in about 200 ms, never writing free text.
Jev Practices & Patterns#community#python#open-models
Community
kshetrajna12/reflexPython
kshetrajna12/reflexPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.