Alibaba Releases Qwen3.8-Max with 2.4 Trillion Parameters; Open Weights May Accelerate Industry Competition

Alibaba launched the Qwen3.8-Max model in August 2026, with total parameters reaching 2.4 trillion and approximately 95 billion active parameters. The model supports a context window of up to 1 million tokens and is now available through the QwenCloud API. Open weights are scheduled for release during the week of August 10 via Hugging Face and ModelScope.

Model Core Specifications and Availability

Qwen3.8-Max is built on the Qwen3.5 architecture and supports text, image, and video inputs with text outputs. The maximum input length is 991,000 tokens, slightly lower in thinking mode; the maximum output is 131,000 tokens, with a reasoning budget cap of 262,000 tokens. API pricing is set at $2 per million input tokens and $6 per million output tokens, with lower costs for cached reads. The model is compatible with OpenAI and DashScope interfaces, requiring only changes to the base URL and model ID for integration.

Benchmark Performance and Real-World Execution

In Terminal-Bench 2.1, Qwen3.8-Max scored 86.6, surpassing Claude Opus 4.8 and Fable 5's 84.6, trailing only GPT-5.6 Sol's 88.8. It scored 67.7 on SWE-bench Pro and 73.5 on FrontierSWE, both below Fable 5's corresponding scores. However, it reached 93.0 on PaperBench, 82.8 on IFBench, and 92.6 on GPQA Diamond. Vision-related benchmarks such as OSWorld-Verified 86.1, Parametric CAD Bench 91.5, and OmniDocBench 1.5 all rank at the top.

Internal tests show the model can independently complete long-horizon tasks. In one autonomous coding demonstration, the model started from an empty repository and ran continuously for over 16 days, generating 265 commits, 127 pull requests, and 151 issues, including automated builds, unit tests, and recovery mechanisms. In another task reproducing a research paper's experiments, the model worked continuously for 125 hours, wrote 7,600 lines of code, completed over 1,100 operations and 33 rounds of GPU training, with the final method improving the AIME24 benchmark by 2.7 percentage points.

Industry Impact of Open Weights

Previously, Qwen series Max-level models did not have open weights. This time, Alibaba plans to release both the 2.4-trillion-parameter version and a 27B small model, allowing users to deploy and fine-tune locally. The 27B version is suitable for standard GPU hardware, while the 2.4-trillion-parameter version still requires a multi-node data center environment. After the release, fields such as software engineering, legal document review, media operations, and design can directly use repository-level coding agents and long-video indexing tools.

The immediate market reaction saw Alibaba's stock rise 7.3% in a single day on the Hong Kong market. Analysts noted that Chinese developers have been able to rapidly narrow the gap with Anthropic and OpenAI despite chip restrictions, and this release serves as another example.

Underlying Drivers

The model has only 95 billion active parameters, with only a portion participating during inference, reducing computational cost and latency. This reflects practical efficiency optimization in the mixture-of-experts architecture, rather than a mere pursuit of total parameter scale. The long-horizon autonomous execution capability stems from built-in support for feedback loops, task allocation, and exception recovery, enabling the model to turn user input into executable projects rather than one-off responses.

Drawing close to or surpassing Fable 5 on some benchmarks shows that Chinese teams' targeted investment in multimodal agents and long-document processing has produced quantifiable results. The open-weights decision likely stems from ecosystem-building considerations—allowing third-party customization to accelerate application deployment while expanding influence under export control constraints.

Independent Assessment

The release of Qwen3.8-Max demonstrates that parameter scale and practical deployment efficiency can coexist. Open weights will accelerate the adoption of Chinese AI applications, but the barrier to running a 2.4-trillion-parameter model locally remains high, and API access will remain the primary usage method in the short term. Its performance in autonomous coding and research reproduction goes beyond single-turn Q&A; its future value depends on whether developers can consistently replicate internal test results in real-world projects.