Alibaba Cloud releases TYQA 2.5, outperforming GPT-4 Turbo
A year after its release, Alibaba’s generic large model TYQA receives a major update.
On May 9th, Alibaba Cloud officially released TYQA version 2.5. According to the parameters disclosed by Alibaba Cloud, TYQA 2.5 has completely surpassed GPT-4 Turbo in model performance. Meanwhile, TYQA’s open-source model with 110 billion parameters has achieved the highest scores on several benchmarks, surpassing Llama-3-70B, and leading the global open-source model arrays.
As of now, the generic large model TYQA has served over 90,000 enterprises through Alibaba Cloud and over 2.2 million enterprises through DingTalk. A large number of SMEs and developers both domestically and internationally are utilizing TYQA by downloading the open-source model, resulting in over 7 million downloads in the open-source communities such as HuggingFace and ModelScope.
With this, the application of TYQA in industries such as PCs, mobile phones, automobiles, aviation, astronomy, mining, education, healthcare, catering, games, and cultural tourism is accelerating.
Matching GPT-4 Turbo
Alibaba Cloud states that compared to the previous version, TYQA 2.5’s understanding, logical reasoning, command compliance, and coding abilities have respectively improved by 9%, 16%, 19%, and 10%. Meanwhile, the results from OpenCompass show that TYQA 2.5 scores equally with GPT-4 Turbo, making it the first domestic large model to achieve this feat.
Corporate clients and developers can integrate TYQA through API calls and model downloads, while individual users can use TYQA for free via the APP, the official website, or the mini-program. On the day of the conference, the original TYQA APP announced its rebranding to “TYQA APP,” intending to become a comprehensive AI assistant that embodies the ability of TYQA large model.
Over 7 million downloads
Last August, TYQA announced their involvement in open-source, and subsequently launched a seamless open-source operation, releasing over a dozen models along the “Full Modality, Full Size” open-source route. At present, the download volume of TYQA open-source models has surpassed 7 million.
The training and iteration costs of large models are extremely high, and most AI developers and SMEs can’t afford them. The open-source trend for large models, driven by Meta, Alibaba Cloud, and others, has greatly accelerated the application of large models by allowing developers to take the initiative in selecting models without having to train them from scratch.
Open-source and openness, becoming the most welcome large model by Chinese enterprises
According to the latest data, TYQA has served over 90,000 enterprises through Alibaba Cloud and over 2.2 million enterprises through DingTalk. It’s now implemented in areas like PCs, mobile phones, automobiles, aviation, astronomy, mining, education, medical, catering, games, and cultural tourism.
On May 9th, Xiaomi’s AI assistant “Xiao Ai Classmate” partnered with Alibaba Cloud’s TYQA large model to enhance its multimodal AI generation capabilities in image generation and image understanding and was implemented in various devices such as Xiaomi cars and mobile phones. Companies like Weibo, Zhongan Insurance, and Perfect World Game also announced their integration with TYQA large model, applying the large model in social media, insurance, and gaming.
In line with Alibaba Cloud’s aspiration to become the “most open cloud in the AI era” by providing an open computing platform, open-source models, and high-quality model services to help customers seize opportunities in the large model era, their open-source strategy is now bringing new business growth for Alibaba Cloud.