On August 3, 2026, the Qwen team released the Qwen 3.8-Max on the official Qwen blog, positioning it as the most powerful model in the current family. The model is already accessible via QwenCloud, but the official plan is to release open weights on Hugging Face and ModelScope next week, so it cannot yet be written as 'Weights Are Downloadable.'
As the scale grows, computing is not fully activated
Qwen3.8-Max continues the architectural foundation of Qwen3.5, with a total of 2.4 trillion parameters and 95 billion parameters activated per inference. The model retains a larger capacity of knowledge and capabilities, while only scheduling some experts to participate in calculations. Please refer to the site's mixed expert model description. However, 95 billion activation parameters still means a considerable reasoning burden; Memory requirements, quantization schemes, license terms, and local deployment costs must be publicly recognized with equal weight and model cards before judgment.
The focus shifted from answering code to completing projects
The official focus is on coding, real work, research, long-range tasks, and multimodal agents as the main upgrade directions. The demonstration included more than ten days of building and self-iterating an engineering project, as well as about 125 hours completing paper reproduction and improvements. The latter has written about 7,600 lines of code, performed over 1,100 operations, and undergone 33 rounds of GPU training. Compared to the early Qwen 3.6-Max-Preview, the focus shifts from single-run programming capabilities to multi-day autonomous execution and feedback loops.
Leaderboard results need to be returned to the test conditions
Qwen's published results show that Qwen 3.8-Max performs well on projects such as PaperBench, OSWorld-Verified, and WideSearch, but the table contains both public and internal benchmarks, and different models may use different proxy frameworks, context lengths, and runtime limits. These numbers are suitable for judging ability directions and should not be taken as absolute rankings in a unified environment. More valuable validation will come from third-party retesting after open weighting, as well as the success rate, recovery capability, and unit task cost in real codebases.
What developers can do now
QwenCloud supports Chat Complications, Responses interfaces compatible with OpenAI specifications, as well as interfaces compatible with Anthropic. The model offers three reasoning_effort levels: xhigh, medium, and low, with xhigh used by default and the thinking state retained by default. At this stage, it is suitable to use fixed task sets to test latency, price, tool call stability, and recovery from long task interruptions; Teams requiring privatization deployment or secondary training should wait until next week to verify weight documents, licenses, and hardware requirements. The real watershed for Qwen 3.8-Max is not the 2.4 trillion figure itself, but whether the open version can bring long-range reliability from cloud demonstrations into a reproducible environment.
