Question 1

Can I really point existing TypeSafe SDK code at this model?

Accepted Answer

Yes for client code. The request and response contract mirrors the TypeSafe POST /v1/systemone shape, so SDK code written for TypeSafe can point at this model with a local endpoint.

Question 2

How is this different from prompting a chat model to decide?

Accepted Answer

It replaces the language-model head with a 256-way slot softmax on a small text tower, continued-pretrained on Thai, and returns calibrated probabilities instead of generated text.

Question 3

Does this run on a laptop CPU, or do I need one of the H100s the configs assume?

Accepted Answer

Inference runs anywhere. The client loads iapp/OpenThai-SystemOne on CUDA, MPS, or CPU, and the bundled server serves the same JSON contract on a port you choose. Retraining is the GPU part: the CPT, SFT, and calibration configs are sized for one H100 at roughly a day, half a day, and hours, and the synthetic-data stage needs an OpenAI-compatible chat endpoint.

Question 4

Where do I check accuracy claims and reproduce the pipeline?

Accepted Answer

The model card with benchmark tables lives on Hugging Face, and the repo ships staged pipeline scripts, configs, and CPU unit tests for the full pretrain to eval flow.

Question 5

What are the pipeline stages, and what does each one cost to run?

Accepted Answer

Eight numbered scripts. Stage 0 does base surgery on CPU and drops the vision encoder. Stage 1 builds 5B Thai-heavy CPT tokens on CPU, then continued-pretrains on 1x H100 for about a day. Stage 2 generates decision and synthetic data on CPU plus an LLM endpoint, then SFT-trains on 1x H100 for roughly half a day. Stage 3 calibrates in hours on 1x H100, stage 4 evaluates on GPU, and stage 5 exports the release to the Hub.

Question 6

Who funded the training, and what is the model built on?

Accepted Answer

Training, evaluation and the free hosted API run on NVIDIA H100 GPUs donated by Siam AI Corporation, which is what makes the open Thai release possible. The base is a Qwen3.5-0.8B text tower whose language-model head is replaced by a 256-way slot softmax, continued-pretrained on Thai, so it produces typed probabilities rather than generated text.

Question 7

What are the practical requirements for inference versus retraining?

Accepted Answer

Inference is light: the pip package plus a model download, and the client loads on CUDA, MPS or CPU, with an HTTP server serving the same JSON contract on any port. Retraining is the heavy side, needing H100-class GPUs per the configs, and the synthetic-data stage additionally needs an OpenAI-compatible chat endpoint configured through SYNTH_BASE_URL, SYNTH_MODEL and SYNTH_API_KEY.

Openthai Systemone

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community