Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

OpenThai-SystemOne

Open Thai and English System One model: Qwen3.5-0.8B with a 256-way slot-softmax head, served with a /v1/systemone-compatible API.

Librarybeta
View on GitHub ↗

Overview

OpenThai-SystemOne is an open Thai and English decision model. It takes a Qwen3.5-0.8B text tower, replaces the language-model head with a 256-way slot softmax, continues pretraining on Thai, and trains it to answer choice, score and noul questions about arbitrary state in one forward pass with calibrated probabilities. The request and response contract mirrors TypeSafe's POST /v1/systemone, so SDK code written for Jev can point at it.

It ships a pip package with typed Choice, Score and Noul classes, a client that runs on CUDA, MPS or CPU, and a uvicorn HTTP server with the same JSON contract. The repository documents the full pipeline: model surgery, 5B Thai-heavy tokens for continued pretraining (about a day on one H100), decision data including synthetic generation, supervised training and calibration, with CPU unit tests. Benchmarks are on the Hugging Face model card. It suits Thai-language applications such as support triage.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich team should handle this? (Thai example)billing · technical · sales
ScoreHow frustrated is the customer?calm · annoyed but polite · very angry
Yes / noIs the customer clearly asking for a refund?yes · no

How it works

  1. pip install openthai-systemone.
  2. Create a SystemOneClient with iapp/OpenThai-SystemOne.
  3. Send state and typed questions, or run the HTTP server.
  4. Use returned probabilities in Thai or English workflows.

Strengths

  • Thai-language focus with English support.
  • Small 0.8B model runs on CPU, MPS or CUDA.
  • Jev-compatible API contract.

Caveats

  • Benchmarks are on the model card, not verified here.
  • Reproducing training needs an H100 and large data.

Getting started

pip install openthai-systemone, then OPENTHAI_SYSTEMONE_MODEL=iapp/OpenThai-SystemOne uvicorn openthai_systemone.server:app --port 8000

At a glance

Type
Library
Status
beta
Language
Python
Runtime
library, FastAPI server
Has code
yes
Has tests
yes

Use cases and integrations

How curated lists describe it

  • OpenThai-SystemOne: open Thai + English System One decision model.

    wh000wh000/awesome-jev-live

  • Thai and English decision model with a slot-softmax head whose request and response shape mirrors the official API, so existing SDK code can point at it.

    MrJev/awesome-jev

Frequently asked questions

Which languages does it support?

Thai and English.

Can I use the TypeSafe SDK with it?

Its server mirrors the /v1/systemone contract, so SDK code can point at it.

How big is it?

About 0.8B parameters with a 256-way slot head.

Related projects

  • Luce

    Open recipe: describe a decision task, have an LLM write training data, and train a Qwen LoRA plus decision head with calibrated outputs.

    ModelbetaPython
  • OpenJev

    OpenJev on Hugging Face: open-weights decision model for text, web pages and screenshots with up to 52 options per forward pass.

    ServicebetaPython
  • WaterSheep

    Apache-2.0 ModernBERT decision model answering Noul, Choice, Score and multi-label questions, with a local Jev-compatible server.

    ServicebetaPython
  • JevBERT

    Unofficial proof of concept serving Jev's /v1/systemone shape from a local BERT-style zero-shot classifier, with honest caveats.

    ServiceprototypePython
  • mini-jev

    Dependency-free Python wrapper that turns any LLM API into a Jev-style choose-one-option decision call for old machines.

    ModelprototypePython
  • d1

    Liquid AI's first decision model for text and images, served at a /v1/systemone endpoint callable with TypeSafe SDKs.

    Libraryproduction