Alibaba launches Qwen3.8-Max, 2.4 trillion‑parameter AI model
On 3 August 2026 Alibaba Group unveiled Qwen3.8‑Max, the company’s largest and most capable artificial‑intelligence model. The system contains 2.4 trillion parameters and uses a mixture‑of‑experts architecture that activates roughly 95 billion parameters for each request, reducing compute cost and latency. Qwen3.8‑Max can handle text, images, video and up to one million tokens in a single operation, enabling analysis of long documents, large code bases and extensive multimedia content.
In internal testing the model autonomously completed a software‑engineering project in 16 days, and on the public Arena AI leaderboard it ranked second worldwide for image‑analysis tasks and among the top Chinese models for text generation, with performance comparable to Anthropic’s Claude Fable 5. Alibaba said the model’s open‑weight files will be released the week of 10 August via Alibaba Cloud’s Model Studio, and it priced token usage at US$2 per million input tokens and US$6 per million output tokens. Following the announcement, Alibaba’s shares rose between 5 % and 7 % on the Hong Kong exchange.
Entities: Alibaba Group Holding · Alibaba Group Holding Ltd · Anthropic · Arena AI · Claude Fable 5 · Hong Kong Stock Exchange · Moonshot AI · Qwen3.8-Max · Qwen3.8‑Max
Claims
What the coverage asserts, and how well corroborated each claim is across sources.
- [● 25 SOURCES] On the Arena AI leaderboard Qwen3.8-Max ranks second globally for image‑analysis tasks, behind Anthropic’s Claude Fable 5. (Alibaba)
- [● 25 SOURCES] Qwen3.8-Max’s benchmark performance is comparable to Anthropic’s Claude Fable 5. (Alibaba)
- [● 25 SOURCES] During internal testing Qwen3.8-Max completed an autonomous software‑engineering project in 16 days. (Alibaba)
- [● 25 SOURCES] Alibaba’s shares rose between 5 % and 7 % on the Hong Kong exchange after the model’s announcement. (Alibaba)
- [● 25 SOURCES] Alibaba released the Qwen3.8-Max model with 2.4 trillion parameters. (Alibaba)
- [● 25 SOURCES] The model uses a mixture‑of‑experts architecture that activates about 95 billion parameters per request. (Alibaba)
- [● 25 SOURCES] Alibaba will release the Qwen3.8-Max model weights for public download in the week starting Aug 10 via Alibaba Cloud Model Studio. (Alibaba)
- [● 25 SOURCES] Qwen3.8-Max can process text, images, video and up to one million tokens per operation. (Alibaba)