Breeze
About
The Breeze large language model (LLM) family, created by MediaTek Research, features a series of open-source models based on the Mistral-7B foundation. Tailored for Traditional Chinese, these models excel in both Traditional Chinese and English languages. They are available in base and instruction-tuned variants, with the latter optimized for tasks like question answering, retrieval augmented generation (RAG), multi-round chat, and summarization. One of Breeze's standout features is its enhanced processing speed, particularly for Traditional Chinese, reaching double the inference speed compared to models like Mistral-7B and Llama 7B. This is largely due to its expanded vocabulary, which includes an additional 30,000 Traditional Chinese tokens. Moreover, the Breeze models perform impressively in benchmarks against other open-source models of similar size 23.