On the occasion of the first anniversary of the release of the Tongyi model, an important historical moment was ushered in. On May 9, Alibaba Cloud officially released Tongyi Qianwen 2.5, and the model performance has fully surpassed GPT-4 Turbo and become the strongest Chinese large model on the surface. At the same time, Tongyi Qianwen's 110 billion parameter open-source model has achieved the best results in multiple benchmarks, surpassing Llama-3-70B and becoming the most powerful model in the open source field.
After more than a year of catching up, the domestic large model has finally entered the core arena and can compete with the first-class foreign large model.
Strive to catch up for a year and achieve the strongest Chinese model
The outbreak of large-scale model technology has been more than a year, and the competition in the industry is fierce and changeable. Since its launch in April 2023, Tongyi Qianwen has always focused on the technical research and development of the basic model, upgrading from the first model to version 2.5. Compared with the previous version 2.1, the comprehension ability, logical reasoning, instruction follower, and code ability of Tongyi Qianwen 2.5 have increased by 9%, 16%, 19%, and 10% respectively, and the Chinese ability continues to lead the industry. According to the evaluation results of the authoritative benchmark OpenCompass, the score of Tongyi Qianwen 2.5 is equal to that of GPT-4 Turbo, which is the first time that this benchmark has recorded such a good result for a domestic large model.
At the same time, Tongyi also released the 110 billion parameter open source model Qwen1.5-110B, which surpassed Meta's Llama-3-70B in MMLU, TheoremQA, GPQA and other benchmark evaluations, and entered the top spot on the HuggingFace Open Source Large Model Leaderboard, once again proving the strongest competitiveness of the Tongyi open source series in the industry.
The general multimodal model and the proprietary capability model also have the top influence in the industry. In a number of multimodal standard tests, the Qwen-VL-Max visual understanding model of Tongyi Qianwen scored better than Gemini Ultra and GPT-4V, and this model has been implemented in many companies. CodeQwen1.5-7B is the top model of Big Code on the HuggingFace code model list, and it is also the base of Tongyi Lingcode, an intelligent coding assistant with the largest number of users in China.
From the former catch-up to today's parallelist, the Tongyi large model frequently dominates the list, which can be said to be the epitome of the domestic large model forging ahead and catching up in the past year.
In the past year, Tongyi has also developed industry-leading capabilities such as Wensheng diagram, intelligent coding, document parsing, audio and video understanding, etc., and enterprise customers and developers can access Tongyi through API calls, model downloads, etc., and individual users can use Tongyi for free from APP, official website and mini program. On the day of the conference, the original Tongyi Qianwen APP announced that it would be renamed "Tongyi APP", integrating the full set of capabilities of the Tongyi large model, and committed to becoming an all-round AI assistant of "Tongyi Dayi".
Firmly open source route and become the strongest open source model in China
In August last year, Tongyi announced that it would join the ranks of open source, and then launched a non-stop open source hurricane, launching more than a dozen models along the "full-modal, full-size" open source route. At present, the number of downloads of the Tongyi open source model has exceeded 7 million.
The training and iteration costs of large models are extremely high, and the vast majority of AI developers and small and medium-sized enterprises cannot afford them. The open-source trend of large models promoted by Meta and Alibaba Cloud allows developers not to train models from scratch, and also gives developers the initiative of model selection, which greatly accelerates the application process of large models.
In order to meet the needs of users in different scenarios, Tongyi has launched eight large language models with parameter sizes ranging from 500 million to 110 billion, such as 0.5B, 1.8B, 4B, 7B, and 14B, which can be easily deployed on mobile phones, PCs and other devices. Large-size models such as 72B and 110B can support enterprise and scientific applications; Medium-sized sizes such as the 32B try to find the best value for money balance between performance, efficiency, and memory footprint. In addition, Tongyi has also open-sourced the visual understanding model Qwen-VL, the audio understanding model Qwen-Audio, the code model CodeQwen1.5-7B, and the hybrid expert model Qwen1.5-MoE.
Tongyi 72B and 110B open source models have topped the Open LLM Leaderboard. On Chatbot Arena, a benchmark platform launched by LMSYS Org, an open research institution, the Tongyi 72B model has repeatedly entered the top 10 in the world in "blind testing" results, creating a precedent for domestic large models.
The long-term precipitation of good reputation has won a lot of fans for the Tongyi open source model, and every time there is an open source action, it will be "squatted" by developers early, and it can always get the first support from ecological partners. "The feedback from developers and the ecological support of the open source community are an important boost to the technological progress of the Tongyi large model." Zhou Jingren, CTO of Alibaba Cloud, revealed that the Tongyi model will continue to be open source in the future.
Open source and openness, to create the most popular large model for Chinese enterprises
Tongyi is becoming the most popular model for Chinese companies. According to the latest data, Tongyi has more than 90,000 enterprises through Alibaba Cloud and more than 2.2 million enterprises through DingTalk, and has now landed in PCs, mobile phones, automobiles, aviation, astronomy, mining, education, medical care, catering, games, cultural tourism and other fields.
On May 9, Xiaomi's artificial intelligence assistant "Xiao Ai" reached a cooperation with Alibaba Cloud Tongyi Model to strengthen its multimodal AI generation capabilities in image generation and image understanding, and landed in Xiaomi cars, mobile phones and other types of devices; Weibo, Zhongan Insurance, Perfect World Games and other companies have also announced access to the Tongyi model, applying the model to social media, insurance, games and other fields.
The Artificial Intelligence Group of the National Astronomical Observatories of the Chinese Academy of Sciences has developed a new generation of astronomical model "Star Language 3.0" based on Tongyi Qianwen, which is the first time that the large model has been applied to the field of astronomical observation. More than 10 mines, including Shaanxi Coal Jianxin Coal Mine, have launched a new major mine risk identification and disposal system supported by Tongyi, which has become the first large-scale implementation of a large model in a mining scenario.
Alibaba Cloud has always emphasized that it will become "the most open cloud in the AI era", and help customers seize the opportunities of the era of large models through open computing platforms, open-source self-developed models, and high-quality model services. Today, the open source strategy is bringing new business growth to Alibaba Cloud.