Alibaba has officially entered the global race for top-tier large language models, intensifying competition even within its own investment portfolio.
On July 19th, Alibaba launched a preview version of its flagship model, Qwen3.8 Max, boasting a massive parameter scale of 2.4 trillion. The company claims its overall capabilities are second only globally to Anthropic's Fable 5. Concurrently, Alibaba has pledged to release the full model weights "soon."
The timing of this release is particularly noteworthy, coming just days after Moonshot AI, a startup backed by Alibaba, stirred global tech stocks with its Kimi K3 model featuring 2.8 trillion parameters, briefly setting a new record for the "largest open model ever." The high-profile debut of Qwen3.8 Max has intensified the internal battle within Alibaba's investment sphere for the title of "largest open model."
The preview version is now available on Alibaba's code development platforms, Qoder and QoderWork, as well as on Token Plan. A public conversational version has also been launched simultaneously.
Preview Version Launched First, Weight Release Timing Uncertain
Alibaba announced this news through an official statement and social media posts from its Tongyi Qianwen team, positioning Qwen3.8 Max as "one of the most powerful models today, comparable to leading frontier AI models."
Developers can now access the preview version early through platforms like Qoder, and a public online chat interface is also open. Alibaba stated it will release the full open weights "soon" but did not provide a specific date.
A developer involved in the project, Bai Shuai, noted that Qwen3.8 Max is the team's first multimodal model to surpass 1 trillion parameters, capable of processing various content formats like images, video, and documents within a single system.
He anticipates this model will comprehensively surpass its predecessor, Qwen 3.7-Max, in complex productivity scenarios such as code generation, full-stack development, data analysis, and office automation.
While the 2.4 trillion parameter count is eye-catching, analysts caution investors and developers to maintain a degree of prudence.
Analysts point out that the total parameter scale does not directly reflect the number of parameters actually engaged in computation per token, making it difficult to estimate true inference costs. To date, Alibaba has not published any benchmark results, nor has it disclosed the activated parameter scale, licensing terms, or a complete model card.
It is worth noting that earlier models in the Qwen series extensively utilized a Mixture-of-Experts (MoE) architecture, which activates only a small fraction of its weights during inference, potentially reducing actual computational power and costs far below the nominal parameter count. If Qwen3.8 Max employs a similar design, the gap between its 2.4 trillion "total parameters" and its actual inference costs could be similarly significant.
Internal Competition Within Alibaba's Investment Sphere
Analysts believe the context of this release reflects profound shifts in the competitive landscape within China's AI ecosystem.
Just days before the Qwen3.8 Max reveal, Moonshot AI, also backed by Alibaba, released Kimi K3. This model, with 2.8 trillion parameters, performed impressively in comparisons with products from Anthropic and OpenAI, briefly causing volatility in global tech stocks and setting a new record for the "largest open model ever."
The emergence of Kimi K3 signals that China's major model labs no longer view open-sourcing weights as a dispensable option but rather as a core bargaining chip in the competition for developer ecosystem dominance. With Alibaba holding stakes in numerous top domestic AI startups, the struggle for the "largest open model" within its ecosystem has effectively evolved into a form of internal competition within its own investment portfolio.
Whether Qwen3.8 Max can fulfill the ambition suggested by its preview version after the full weights are released remains to be seen through benchmark testing and market validation.
Comments