Is Bekko System One the same thing as TypeSafe's Jev?

No. It is an independent experimental project in the same System One category, reimplementing Choice, Noul, and Score decisions with open weights from 17M to 400M parameters; the README compares its v0 results with Jev 1.13.

Can Bekko models run without sending data to a server?

Yes. Inference is available as standalone Python on CPU or CUDA and as ONNX in Node.js or the browser; the browser app states that input text is not sent to an inference server.

How do the models compare with Jev on the benchmark?

On the project's S1MB snapshot dated 2026-09-30, which spans 137 benchmarks, 106 subsets, 14,009 cases and 26,269 decisions, Bekko 400M scores 50.60 Task Average against Jev 1.13's 59.59. The gap widens on adaptation tasks, 54.48 against 96.27, and Jev's parameter count is not disclosed in that snapshot.

What does the environment require?

Clone the repository and run `uv sync --locked`. The training environment targets Linux x86_64 with Python 3.12, and the lockfile pins PyTorch 2.10.0 with CUDA 13.0, Transformers 5.17.0 and Sentence Transformers 6.1.0. CPU tests and SDPA inference do not need FlashAttention; GPU training with the shipped configs needs `uv sync --locked --extra fa2` and a compatible NVIDIA driver.

What does a prediction actually return?

A result keyed by decision id, with selected_id for the chosen candidate, probability_yes for an authored true condition, and a score on the scale you supplied, plus probabilities keyed by candidate id. Passing a list processes several requests and returns results in the same order, with batch_size bounding the window and token_budget setting estimated work per microbatch rather than a hard memory limit. Probabilities do not establish calibrated confidence.

How do I train it on my own data?

Start from the reranker or typed-decisions config, copy it, set your data paths and a new output directory. Each JSONL row describes one complete candidate group with a query, candidates, a task tag and a target probability distribution in candidate order; the files under data are placeholders. Full fine-tuning, LoRA and soft targets are supported, and the two templates train separate models with their own weights.

What licenses cover the code, weights and dataset?

The code is MIT, but the model weights and the training dataset carry their own licenses on Hugging Face, so check each model card. The v0 family has limited generalization, and because the training set shares task families with the evaluation, strong S1MB results do not establish broad generalization.

CommunityJev Models & Reimplementations

Bekko System One

Independent training toolkit and open weights for System One decision models in the same category as Jev: 17M, 68M, and 400M models answering Choice, Noul, Score, and reranking questions, with Python and ONNX inference and a documented generalization gap to Jev 1.13 on the project's S1MB benchmark.

See highlights

Open github.com

Stars: 36
Language: Python
License: MIT
Published by: Community
Last verified: Oct 2026

hotchpotch/bekko-system-one

User repository on GitHub

View on GitHub

Small System One decision models (17M–400M) for Yes/No, Choice, Score, and reranking. Training toolkit, Python inference, and ONNX support for Node.js and browsers.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Release article · Models · Quickstart · Benchmarks · Browser app · Training guide · Training dataset. Bekko System One is an experimental project for training and running very small System One decision models, in the same category as TypeSafe AI's Jev. The v0 family spans 17M to 400M parameters. These models turn instructions, context, and candidates into typed decisions and probabilities that software can use directly.

Quoted from the project's README. Read the full README.

Tagscommunity decision-models onnx weights pytorch

Works forJev for Classification Jev for Scoring Jev for Ranking Jev for Verification

Used inJev for Software & DevTools

Languages: Python
Platforms: Browser
Capabilities: Open sourceRuns locally

Highlights

The v0 family spans 17M, 68M, and 400M parameters, exported as standalone Python and ONNX for Node.js and browser inference.
On the project's S1MB benchmark, Bekko 400M scores 50.60 Task Avg versus Jev 1.13's 59.59; on adaptation tasks the gap is 54.48 versus 96.27.
Training supports full fine-tuning, LoRA, and soft targets on a public Hugging Face dataset, with configs for each model size.
The browser app runs Choice, Noul, and Score examples on-device with CPU or WebGPU; input text is not sent to an inference server.
The quoted S1MB snapshot covers 137 benchmarks, 106 subsets, 14,009 cases, and 26,269 decisions as of 2026-09-30.

Watch out

Code is MIT, but model weights and the training dataset carry their own licenses on Hugging Face; v0 has limited generalization and shares task families with its evaluation set, so S1MB scores are not raw accuracy.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

AutoJev

130

An open recipe for a 27B multimodal decision model trained with full-weight SFT on 73,000 examples over one H200, serving choice, noul and score on /v1/systemone with per-checkpoint provenance and released calibration plots.

Jev Models & Reimplementations#community#calibration#qwen

Community

denis-pplx/autojevPython

RSI-Jev

109

Open decision models from a self-improvement loop that publishes every experiment, failures included. They answer Noul, Choice and Score questions on the same POST /v1/systemone wire format as Jev, returning a calibrated probability per option in one forward pass; the 27B scores 65.64 on the Decision Index 0.3 set.

Jev Models & Reimplementations#community#decision-models#system-one

Community

Shanghua-Gao/RSI-JevHTML

Vev

Jev-compatible decision models with vision input: LoRA fine-tunes of Qwen3.5 at 4B and 9B serving Choice, Score and Noul on the /v1/systemone wire format, with screenshots and photos placed beside text in the state. Runs on your own NVIDIA GPU; weights are non-commercial.

Jev Models & Reimplementations#community#vision#open-weights

Community

Xiaooolong/vevPython

WaterSheep

Apache-2.0 ModernBERT fine-tune that answers noul, choice, score and multi-label questions with a probability for every option. It serves Jev's POST /v1/systemone locally so TypeSafe's Python SDK runs unchanged, reporting 0.026 expected calibration error on its test split.

Jev Models & Reimplementations#community#calibration#modernbert

Community

SamratDuttaOfficial/WaterSheepPython

ruling

A local Jev-compatible server that reads Choice, Score and Noul answers from the logits of any MLX checkpoint or OpenAI-compatible endpoint with no training, serving POST /v1/systemone as a drop-in for the official TypeSafe SDK.

Jev Models & Reimplementations#community#mlx#calibration

Community

bradAGI/rulingPython

Visual-Jev

A Jev-style typed-decision baseline built on the frozen Qwen3.5-4B checkpoint that accepts a local image path directly, letting the vision-language model score declared options without generating text. Ships scoring modes, a WebGPU browser demo and benchmark results measured on an RTX 3090.

Jev Models & Reimplementations#community#multimodal#open-models

Community

jiangxiluning/Visual-JevPython

From the Jev guides

Original write-ups on what this entry is for.

Multimodal Decision Models: What Clef and d1 Change

Oct 12, 2026

Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.

Jev Models & Reimplementations#multimodal#open-models#decision-model

Read article

The Decision-Model Price War

Oct 11, 2026

One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.

Jev Models & Reimplementations#decision-model#benchmarks#pricing

Read article

What Is Actually on Hugging Face

Oct 8, 2026

Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.

Jev Models & Reimplementations#open-models#hugging-face#local

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai