Alibaba unveils Qwen3.8, the massive AI model with 2.4 trillion parameters
The Qwen team at Alibaba has presented Qwen3.8, a multimodal AI model with 2.4 trillion parameters that the company describes as one of the most powerful systems available today. In the announcement, Alibaba stated that the model is second only to Fable 5, the flagship system from Anthropic, but did not publish any independent benchmarks to support this comparison. Qwen3.8 is immediately available in preview version via Token Plan, Qoder, and QoderWork, for one-tenth of the standard price.
Qwen3.8 is launching and going open-weight soon!? With a massive 2.4T parameters, this model is continuously evolving. We believe it's one of the most powerful models available today, compatible with leading frontier AI models, second only to Fable 5.
You don't have to wait to pic.twitter.com/JS3ID73IYS – Qwen (@Alibaba_Qwen) July 19, 2026
The new model adopts a sparse Mixture-of-Experts architecture and represents the first multimodal system from the Qwen team to exceed the trillion-parameter threshold. It manages text, images, video, and documents in a single processing stream. Alibaba has not disclosed how many of the 2.4 trillion parameters are actually active during inference, a crucial detail for sparse MoE models where only a fraction of the total is engaged with each request.
The launch comes amid an escalation in the race for parameters among Chinese labs. Moonshot AI, in which Alibaba controls about 36% of the shares, has recently released Kimi K3 with 2.8 trillion parameters, briefly becoming the largest open-source model in the world. Thus, Qwen3.8 ranks as the second largest publicly known model, following Kimi K3, while Zhipu AI is also moving forward with GLM 5.2.
Alibaba's AI strategy does not stop with the Qwen family. The group has recently secured the role of technology partner for Apple Intelligence in China, after Beijing authorities approved the service. This agreement opens a direct channel for Alibaba to Apple's vast user base in the country, a commercial advantage that adds to the range of Qwen models.
Pricing, compatibility, and the issue of open weights Access to the preview is granted through Token Plan subscriptions. The Lite plan costs $6 per week for 2,500 credits, while the Pro plan increases to $68 for 40,000 weekly credits and allows running 6 to 8 agents in parallel. Both levels ensure compatibility with OpenAI and Anthropic API protocols, so developers can integrate Qwen3.8 without rewriting existing pipelines. Alibaba has not released a pricing list for standalone API access.
The predecessor, Qwen3.7-Max, achieved 92.4% on the GPQA Diamond benchmark, 80.4% on SWE-bench Verified, and 69.7% on Terminal-Bench 2.0, with a cost of $1.25 per million tokens in input and $3.75 per million in output. Alibaba has promised an open-weight version of Qwen3.8 soon, without indicating a date. This would mark a significant shift from established practice, as Max-tier models have always remained proprietary.
Alibaba's shares closed at $114.97 on July 17, down 2.14%, for a market capitalization of about $262 billion. The announcement came on Saturday, after U.S. markets closed; any effects on the stock will only be seen when trading reopens on Monday. Without verifiable benchmarks, the claim of having nearly reached Fable 5 remains a statement yet to be proven.