Blab Tech
Alibaba Unveils 2.4-Trillion-Parameter Qwen3.8-Max, Matching Leading U.S. Models on Benchmarks
Alibaba Group announced the launch of Qwen3.8‑Max, a multimodal large‑language model that it claims is the most powerful model it has released to date. The model contains 2.4 trillion parameters and is built on a mixture‑of‑experts architecture, with only about 95 billion parameters active for any given task. Alibaba said the model can handle up to a 1‑million‑token context window and can perform text, image, video, and document tasks.
In a demonstration released with the announcement, Qwen3.8‑Max was tasked with reproducing an experiment described in a research paper. The model was given only the paper and, over a period of roughly 125 hours—just over five days—generated all of the code, ran the experiment, and produced a written report. Alibaba described the exercise as an example of the model’s ability to carry out long‑horizon, low‑human‑involvement tasks.
Benchmark results posted by Alibaba show that Qwen3.8‑Max is on par with OpenAI’s GPT‑5.6 Sol and Anthropic’s Fable 5 on several standard tests. The model also outperformed both competitors on two specific evaluations: a visual‑reasoning benchmark and a test of agentic computer use. According to the company’s data, the model’s scores on these tests are higher than those reported for GPT‑5.6 Sol and Fable 5.
Alibaba said the model’s weights will be released to the public next week. Pricing for the model’s API is set at $2 per million input tokens and $6 per million output tokens, which is lower than the $10 and $50 rates that Anthropic charges for Fable 5.
The announcement comes amid a broader trend of Chinese AI developers releasing models that approach the performance of the most advanced U.S. offerings. Last month, Moonshot AI, a Beijing‑based company, introduced Kimi K3, which it said performed close to the top U.S. models in published tests. At the same time, a growing number of U.S. users have switched from OpenAI and Anthropic subscriptions to cheaper Chinese alternatives.
Public sentiment about AI also differs markedly between the United States and China. A 2023 poll conducted by KPMG International and the University of Queensland found that 95 % of Chinese respondents expressed optimism about AI, while only 41 % of U.S. respondents felt that the benefits outweighed the risks. The poll also reported that 36 % of U.S. respondents were fearful of AI.
Alibaba’s marketing materials for Qwen3.8‑Max contrast with recent U.S. AI advertisements. Alibaba’s promotional video shows the model working on long‑running tasks while people engage in leisure activities, a tone that echoes the ambient “lofi beats” videos common on YouTube. In comparison, Anthropic’s recent commercial, which aired during the World Cup, featured dramatic imagery and a narrative aimed at reassuring viewers about AI safety.
The release of Qwen3.8‑Max adds to the growing number of large‑language models that are available under open‑source or open‑weights licenses. Alibaba’s Qwen family has historically distributed many models under the Apache License or a non‑commercial research license, and the company has stated that it will continue to provide open access to its models.
At present, the U.S. government has not announced new regulatory actions specifically targeting Qwen3.8‑Max or other Chinese models. However, the increasing parity between Chinese and U.S. models has intensified discussions about national security, export controls, and the potential need for new policy frameworks.
In summary, Alibaba’s Qwen3.8‑Max represents a significant technical milestone for the company and for the broader Chinese AI ecosystem. The model’s performance on key benchmarks, its lower pricing, and its upcoming open‑weights release are likely to influence the competitive dynamics of the global AI market. The next few weeks will see developers testing the model’s capabilities and the industry monitoring how the U.S. regulatory environment responds to the continued convergence of AI technology across borders.
In a demonstration released with the announcement, Qwen3.8‑Max was tasked with reproducing an experiment described in a research paper. The model was given only the paper and, over a period of roughly 125 hours—just over five days—generated all of the code, ran the experiment, and produced a written report. Alibaba described the exercise as an example of the model’s ability to carry out long‑horizon, low‑human‑involvement tasks.
Benchmark results posted by Alibaba show that Qwen3.8‑Max is on par with OpenAI’s GPT‑5.6 Sol and Anthropic’s Fable 5 on several standard tests. The model also outperformed both competitors on two specific evaluations: a visual‑reasoning benchmark and a test of agentic computer use. According to the company’s data, the model’s scores on these tests are higher than those reported for GPT‑5.6 Sol and Fable 5.
Alibaba said the model’s weights will be released to the public next week. Pricing for the model’s API is set at $2 per million input tokens and $6 per million output tokens, which is lower than the $10 and $50 rates that Anthropic charges for Fable 5.
The announcement comes amid a broader trend of Chinese AI developers releasing models that approach the performance of the most advanced U.S. offerings. Last month, Moonshot AI, a Beijing‑based company, introduced Kimi K3, which it said performed close to the top U.S. models in published tests. At the same time, a growing number of U.S. users have switched from OpenAI and Anthropic subscriptions to cheaper Chinese alternatives.
Public sentiment about AI also differs markedly between the United States and China. A 2023 poll conducted by KPMG International and the University of Queensland found that 95 % of Chinese respondents expressed optimism about AI, while only 41 % of U.S. respondents felt that the benefits outweighed the risks. The poll also reported that 36 % of U.S. respondents were fearful of AI.
Alibaba’s marketing materials for Qwen3.8‑Max contrast with recent U.S. AI advertisements. Alibaba’s promotional video shows the model working on long‑running tasks while people engage in leisure activities, a tone that echoes the ambient “lofi beats” videos common on YouTube. In comparison, Anthropic’s recent commercial, which aired during the World Cup, featured dramatic imagery and a narrative aimed at reassuring viewers about AI safety.
The release of Qwen3.8‑Max adds to the growing number of large‑language models that are available under open‑source or open‑weights licenses. Alibaba’s Qwen family has historically distributed many models under the Apache License or a non‑commercial research license, and the company has stated that it will continue to provide open access to its models.
At present, the U.S. government has not announced new regulatory actions specifically targeting Qwen3.8‑Max or other Chinese models. However, the increasing parity between Chinese and U.S. models has intensified discussions about national security, export controls, and the potential need for new policy frameworks.
In summary, Alibaba’s Qwen3.8‑Max represents a significant technical milestone for the company and for the broader Chinese AI ecosystem. The model’s performance on key benchmarks, its lower pricing, and its upcoming open‑weights release are likely to influence the competitive dynamics of the global AI market. The next few weeks will see developers testing the model’s capabilities and the industry monitoring how the U.S. regulatory environment responds to the continued convergence of AI technology across borders.
