Bekko System One
Experimental 17M-400M English decision models for choice, yes/no, score and reranking, with weights, training code and an ONNX browser demo.
Overview
Bekko System One is an experimental project for training and running very small System One decision models in the same category as Jev. The v0 family spans 17M to 400M parameters and turns instructions, context and candidates into typed decisions with probabilities: Choice returns a candidate ID, Noul a yes/no probability, Score a rating on your own scale, plus reranking of retrieved passages.
It provides training and inference with Sentence Transformers, standalone Python and ONNX exports, a training dataset and guide, and a browser app with editable questions. The README is frank that v0 has limited generalization: some benchmark results are strong, but instruction- and context-adaptation results are weaker, and its training set overlaps dataset families in the evaluation. List summaries note it remains substantially behind Jev on the project's generalization tests. It introduces the System One Mosaic Benchmark (S1MB).
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does the customer ask for a refund? | yes · no |
| Choice | Which team should handle this support request?(inferred) | — |
| Score | Rate this answer on your own scale | — |
How it works
- Pick a model size
- Pass instructions, context and candidates
- Get typed decisions with probabilities
- Deploy via Python, ONNX or the browser
Strengths
- Very small models run locally or in a browser
- Open weights, data and training code
- Honest about limits
Caveats
- Limited generalization in v0
- English only
- Train/eval overlap noted by the author
Getting started
See the Quickstart in the README and the release article on Hugging Face.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- Sentence Transformers / ONNX / browser
- Has code
- yes
Use cases and integrations
How curated lists describe it
Open decision models: independent 17M–400M English models for Choice, Noul, and Score, with public weights, training code, and an [ONNX browser demo](https://huggingface.co/spaces/hotchpotch/bekko-system-one-in-browser); v0 remains substantially behind Jev on the project's generalization tests.
yibie/awesome-jevSmall System One decision models (17M–400M) for Yes/No, Choice, Score, and…
wh000wh000/awesome-jev-liveIndependent 17M–400M English decision models for Choice, Noul, and Score, with public weights, training code, and an [ONNX browser demo](https://huggingface.co/spaces/hotchpotch/bekko-system-one-in-browser); v0 remains substantially behind Jev on the project's generalization tests.
cobanov/awesome-jevSmall decision models from 17M to 400M for yes/no, choice, score and reranking, with a training toolkit, Python inference and a browser app.
MrJev/awesome-jev
Frequently asked questions
How big are the models?
17M to 400M parameters.
How does it compare to Jev?
The project's generalization tests show v0 substantially behind Jev.
Can it run in a browser?
Yes, via ONNX in the browser demo.
Related projects
AutoTrust's Apache-2.0 student of Jev 1.13 on Qwen3.8-27B, serving typed decisions and ordinary generation from one vLLM engine.
ModelbetaPythonMangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++