Alibaba unveils Qwen3.8-Max, a 2.4-trillion-parameter AI model that rivals Anthropic and OpenAI on key benchmarks

Alibaba on Monday unveiled Qwen3.8-Max, describing it as its largest and most capable artificial-intelligence model to date. The system carries 2.4 trillion parameters, uses a mixture-of-experts architecture that activates only 95 billion parameters per request, and can process up to 1 million tokens at a time. Alibaba said it will release the model's weights next week through Alibaba Cloud's Model Studio platform, marking a return to open-weight releases after a brief pivot toward proprietary models earlier this year.

Benchmark positioning

On the crowdsourced comparison platform Arena.AI, Qwen3.8-Max immediately became the highest-ranking Chinese model for text tasks, though it still trails Anthropic's Claude Fable 5 and three Opus variants. For visual analysis, it ranked second globally, behind only a Fable 5 variant. In front-end coding, it was beaten only by two Opus models and Moonshot AI's Kimi K3. Alibaba's own internal testing showed the model broadly matching, and sometimes exceeding, Fable 5 and OpenAI's GPT-5.6 Sol across several benchmarks, surpassing both on visual reasoning and agentic computer use.

Autonomous task execution

Alibaba highlighted the model's ability to handle long-horizon agentic tasks with minimal human involvement. In one internal test, Qwen3.8-Max worked autonomously for around 125 hours (close to five days) to replicate an experiment described in a research paper. Given only the paper itself, the model created all the necessary code from scratch, analyzed the data, ran the experiment, and reported the results. Alibaba characterized this as exactly the type of work that takes skilled engineers days. The company also said the model completed a software-engineering project in 16 days.

Competitive landscape

Qwen3.8-Max enters a field where Chinese labs are rapidly closing the gap with US frontier developers. Moonshot AI launched its Kimi K3 model last month with 2.8 trillion parameters, and published test results indicating it too trailed closely behind the most advanced American-made systems. Unlike OpenAI, Anthropic, and Google, which do not disclose parameter counts for their closed-source models, Chinese firms routinely publish these figures to gain traction among developers. Open-weight releases have become a norm in China's AI industry and a growing point of differentiation.

Pricing and market reaction

Alibaba priced Qwen3.8-Max at $2 per million input tokens and $6 per million output tokens, sharply undercutting Anthropic's Fable 5, which costs $10 per million input tokens and $50 per million output tokens. US-listed Alibaba shares rose around 3.5 percent in pre-market trading following the announcement.

Geopolitical context

The launch adds to tensions in Silicon Valley and Washington over how to retain the US technological edge. US export restrictions on advanced chips and semiconductor manufacturing equipment, begun in 2022 and later tightened, were intended to slow China's AI progress. Several analysts noted that the restrictions have instead pushed China to reduce its dependence on US technology, with models like Qwen3.8-Max demonstrating that the gap has narrowed considerably.

Qwen3.8-Max (Alibaba)2.4

Kimi K3 (Moonshot AI)2.8

Qwen3.8-Max (Alibaba)

2.4 trillion parameters

Kimi K3 (Moonshot AI)

2.8 trillion parameters

Data from the article text

Qwen3.8-Max input2

Qwen3.8-Max output6

Fable 5 input10

Fable 5 output50

Qwen3.8-Max input

2 $

Qwen3.8-Max output

6 $

Fable 5 input

10 $

Fable 5 output

50 $

Data from the article text