OpenJAI-v1.0-14B โมเดลไทยจากทีมเดียวกับ JaiTTS ที่เปิดฟรี

OpenJAI-v1.0-14B is an open-source large language model developed by a team working alongside JaiTTS, which is now freely available. Created by Nokka, the model was trained on Qwen3-14B and focuses on three key areas: instruction following, long-context understanding, and tool calling. The team claims that the model's knowledge base remains intact after fine-tuning, avoiding what's called catastrophic forgetting.

The model's performance has been documented on arXiv as "OpenJAI-v1.0: An Open Thai Large Language Model." In a comparison table, OpenJAI-14B outperforms Qwen3-14B in seven out of ten benchmark tests, while falling short in two areas: MMLU-ProX-lite-EN and MMLU-ProX-lite-TH. Despite this, the team emphasizes that the model's knowledge base remains stable.

The model can be run on a single machine, with a maximum input length of 32,768 tokens. It supports open-source licensing under Apache 2.0 and has gained 589 downloads on Hugging Face. The team encourages users to try the model for tasks that require function calling or model chaining, particularly in the Thai language domain.