Alibaba introduces its new AI model, claims it surpasses DeepSeek-V3

Alibaba announced a new version of its Qwen 2.5 artificial intelligence model on Wednesday, the first day of the Lunar Year in China. The Chinese tech company argued that Qwen 2.5 surpassed the highly-acclaimed DeepSeek-V3. 

Alibaba posted on its official WeChat account that Qwen 2.5-Max outperformed almost all AI models, including GPT-40, DeepSeek-V3, and Llama-3.1-405B. The company’s news came in the wake of the release of DeepSeek’s AI assistant (DeepSeek-V3) on January 10 and the January 20 release of its R1 model. 

The recent release of DeepSeek’s R1 model sent a shockwave in the U.S. tech companies’ stocks tumbling, especially Nvidia. DeepSeek claimed that it created the R1 model with only $6 billion compared to the billions of dollars other U.S. tech firms spend in the AI space. 

The tech company also caused some momentum in its own domestic market, with other Chinese tech firms rushing to release updates to their AI models. A report from Reuters revealed that two days after the release of DeepSeek-R1, ByteDance also released an update to its flagship AI model. The TikTok parent company argued that it outperforms Microsoft-backed OpenAI’s o1 in tests that measure how well AI models understood and responded to complex instructions.

Alibaba releases its new version of the Qwen2.5 AI model

The burst of DeepSeek V3 has attracted attention from the whole AI community to large-scale MoE models. Concurrently, we have been building Qwen2.5-Max, a large MoE LLM pretrained on massive data and post-trained with curated SFT and RLHF recipes. It achieves competitive… pic.twitter.com/oHVl16vfje

— Qwen (@Alibaba_Qwen) January 28, 2025

The Chinese tech company announced on Monday that the new AI model, Qwen2.5-VL, could perform a number of text and image analysis tasks. The firm also said Qwen2.5 was similar to the model powering OpenAI’s recently launched Operator. The AI model can understand videos, parse files, and count objects in images, as well as control a PC.

According to benchmarking tests conducted by the Qwen team, the Qwen2.5-VL model outperforms OpenAI’s GPT-40, Anthropic’s Claude 3.5 sonnet, and Google’s Gemini 2.0 Flash. The new AI model could beat its rivals in video understanding, math, document analysis, and question-answer evaluations.

Alibaba confirmed that the Qwen2.5-VL was available for testing in its Qwen Chat app and for downloading from the AI dev platform Hugging Face. The Qwen team said that the AI model can analyze charts and graphics, extract data from scans of invoices and forms, and “comprehend” multiple-hour-long videos. The AI model can also recognize IPs from films and TV series, as well as a wide variety of products.

The Qwen team disclosed that the model had certain restrictions on topics it could discuss in Qwen Chat due to the fact that the AI was developed by a Chinese company. According to the team, China’s internet regulator gauges many models developed in the country to ensure their responses “embody core socialist values.” Several Chinese AI companies, like Ernie, also deflect responses to topics that might raise the ire of regulators or that might be deemed too sensitive. 

Qwen’s team reveals Qwen2.5-VL’s capabilities

Qwen2.5-VL’s dev team revealed that one of the AI model’s interesting features is its ability to interact with software, both on PCs and mobile devices. Philipp Schmid, a technical lead at Hugging Face, showed the AI model launching the Booking.com app for Android and booking a flight from Chongqing to Beijing. 

“Despite all the DeepSeek Hype, Qwenn just dropped the best open Multimodal! Qwen 2.5 VL is a Vision Language Model that can control your computer, similar to the OpenAI operator, extract structured information from charts, and more!!”

– Philipp Schmid, Tech Lead at Hugging Face

Vaibhav Srivastav, data scientist at Hugging Face, showed how the Qwen2.5-VL model controls apps on a Linux desktop but couldn’t accomplish much beyond switching tabs. The demonstration aligned with Qwen’s benchmarking, which showed Qwen2.5-VL scored poorly on OSWorld, a benchmark that tries to mimic a real computer environment.

The Chinese AI tech company also revealed that the two smaller, less sophisticated models in the Qwen2.5VL series, Qwen2.5-VL-3B and Qwen2.5-VL-7B, were available under permissive licenses. The flagship Qwen2.5-VL-7B will still be under Alibaba’s custom license, which requires that firms and devs with more than 100 million monthly active users request permission from Qwen or Alibaba before deploying the AI model commercially.

Cryptopolitan Academy: How to Write a Web3 Resume That Lands Interviews - FREE Cheat Sheet