- Superior performance against current open-source Thai language models, comparable to GPT-3.5.
- The Typhoon-7B pre-trained model is open source and freely available for download to improve Thailand’s AI industry’s competitiveness.
- The Typhoon instruction-tuned model is available through an API service.
SCB 10X, a subsidiary of SCBX Group, is launching an innovative Large Language Model optimized for Thai. Currently, it is the leading Thai large language model that performs comparably to GPT-3.5. Its evaluation is based on Thai language exams of equivalent difficulty levels to high school and standardized exams in Thailand. “Typhoon” addresses the language gap prevalent in models primarily trained in English by overcoming resource limitations associated with the Thai language.
Two versions of Typhoon are available: the Pretrained Model, which imparts Thai language proficiency to the model, encompassing vocabulary, context, cultural nuances, and global knowledge, and the Instruction-tuned Model, which extends the model’s ability to execute translations, summaries, and questions based on its understanding from pretraining. Compared to publicly available Thai language models, Typhoon’s experimental results demonstrate its superiority. Thai text can be processed 2.62 times faster than text processed by GPT-3.5 despite having only 7 billion parameters.
The Thai Large Language Model represents a major step forward in AI language models for Thai, which opens up several new applications in linguistics. It underlines SCB 10X’s commitment to advancing technology within the SCBX Group. In addition to fostering SCB 10X Group, it aims to become an AI-first organization through innovative service solutions and advanced technology. “Typhoon”, the Pre-trained Model version, is now available for developers and interested parties to download for free via (https://huggingface.co/scb10x/typhoon-7b) under the Apache 2.0 license. More technical information is available at https://arxiv.org/abs/2312.13951. Additionally, those interested can soon register to test the initial version of the Instruction-tuned Model through an API page link (https://opentyphoon.ai), contributing to the development of the Thai Large Model (Thai LLM) for increased efficiency and advancement, ultimately enhancing the Thai AI industry’s competitiveness.