Shisa V2.1 Series Released: Supporting Everything from Devices to the Cloud
Introducing a more efficient and powerful LLM series with significantly improved Japanese performance.
Shisa.AI is proud to release the "Shisa V2.1" series, a bilingual general-purpose chat model (press release). Building upon the highly acclaimed Shisa V2, this new series delivers enhanced Japanese language performance and superior efficiency.
The series is available in five sizes, ranging from a 1.2B model optimized for mobile devices to 70B models that rival GPT-4 in Japanese proficiency. Beyond free access on Hugging Face, we are also pleased to introduce our new paid API service.
These improvements were achieved by overhauling over 80% of our training datasets and incorporating advanced techniques like model merging alongside traditional SFT and DPO training.
More from the newsroom
Shisa V2 Series Released: Next-Generation Bilingual LLMs
Introducing our latest bilingual LLM series featuring significant advancements in Japanese language processing.
Read moreShisa 7B released
A bilingual general-purpose chat model using a synthetic-data driven approach.
Read more