Alibaba Open-Sources 2.4 Trillion Parameter AI Model Qwen3.8

Alibaba Cloud uploaded the full open weights for its largest artificial intelligence system, Qwen3.8-2.4T-A95B, to Hugging Face and ModelScope. The release gives developers direct access to the files of the flagship architecture.

Credit: Qwen.ai

The model contains 2.4 trillion total parameters. It runs on a sparse mixture-of-experts design that activates 95 billion parameters during each processing step.

System specifications list a native context capacity of 262,144 tokens, which expands up to 1.01 million tokens. Output buffers allow generation up to 128,000 tokens for handling software code and long document sets.

Alibaba described the architectural scale in official documentation:

"Built upon the architectural foundation of Qwen3.5, Qwen3.8-Max scales to 2.4 trillion parameters, delivering comprehensive improvements across coding, work, research and long-horizon tasks."

The open-weight build operates with internal reasoning steps turned on by default. Users can adjust reasoning intensity through a setting called reasoning effort, which accepts high, medium, and low parameters.

The downloadable distribution differs from the cloud-hosted API version. The open release omits direct image inputs and disables the option to turn off reasoning steps entirely.

Commercial licensing terms set a distinct rule based on company size. Organizations generating under $50 million in annual revenue can download and run the model files free of charge.

Businesses exceeding $50 million in yearly revenue must secure a commercial agreement with Alibaba Cloud before using the weights in commercial products or serving tools.

Hardware requirements for the 2.4 trillion parameter files demand data-center infrastructure. Hardware partner NVIDIA published early performance figures following test runs on specialized compute clusters.

"Alibaba released the open weights for Qwen3.8-2.4T-A95B (Qwen3.8-Max), its largest open-weight model, bringing near-frontier capabilities to the open ecosystem."

NVIDIA reported processing speeds exceeding 4,000 tokens per second per GPU on GB300 NVL72 hardware setups in 8-bit precision. Third-party software teams, including Unsloth AI, released compressed 1-bit versions that shrink the original 4.9-terabyte file size down to 397 gigabytes.

Internal evaluation tests show the model running long-horizon autonomous tasks over extended periods. In one test run, the system operated continuously for 125 hours, generating 7,600 lines of code across 33 GPU training cycles to reproduce scientific research.

In a separate business simulation, the model managed simulated online retail stores, quadrupled its starting capital, and flagged fraudulent vendors without human assistance.

The release follows recent open model announcements from rival firm Moonshot AI, which debuted its Kimi K3 architecture weeks earlier.

Alibaba confirmed plans for a follow-up release named Qwen3.8-27B. That smaller 27-billion parameter version will target software developers operating standard desktop computers.