Qwen 3.8-Max - A flagship large-scale model launched by the Alibaba Cloud Qwen team.

What is Qwen 3.8-Max?

Qwen 3.8-Max is the flagship large-scale model launched by the Alibaba Cloud Qwen team, with a total of 2.4 trillion parameters, and is an extension of the Qwen 3.5 architecture. As the most powerful model in the Qwen family to date, it is the first Max-level model with weights to be open-sourced. The model has achieved comprehensive upgrades in programming, office work, scientific research, and long-term tasks, supports calls through the Qianwen AI platform API, is compatible with the OpenAI and Anthropic protocols, and can be seamlessly integrated into mainstream intelligent agent frameworks such as Claude Code, Codex, and Qoder.

Main functions of Qwen 3.8-Max

• Long-range automatic programmingSupports autonomous programming within ten days, enabling the construction of a self-evolving harness and the completion of requirement collection, issue dispatch, code generation, and self-repair.
• Paper Reproduction and SurpassingAfter independently reproducing the paper's experiments, 18 improvement ideas were tested through four rounds of self-evolutionary cycles, resulting in a 2.7-point improvement over the original method on AIME24.
• Competition PracticeWithin 24 hours, the solution was independently built and iterated, and the accuracy rate increased from 0.60 to 0.853, beating 87% of the human teams.
• Improving office efficiencyIt covers hundreds of high-value professional scenarios, such as lawyers identifying 1,284 clauses in one hour and designers creating an 8-page interactive prototype in one go.
• Quantitative Strategy DevelopmentStarting from a single sentence description, approximately 330 sub-agents are scheduled, and about 6,000 backtests are completed to deliver an end-to-end ETF rotation strategy.
• Independent chip designAfter approximately 500 rounds of interaction, the number of gates in the cryptographic accelerator netlist was optimized from 8,298 to 678, the area was reduced by 81%, and temporal closure was achieved.

Technical Principles of Qwen 3.8-Max

• Architecture extensionIt is an extension based on the Qwen 3.5 architecture, with a total of 2.4 trillion parameters and 95B activated parameters, achieving a leap in scale while maintaining architectural inheritance.
• Real-world reinforcement learning systemsBy jointly expanding the number of RL environments and training computing power, we enhance general-purpose capabilities and address three core coupling challenges: decoupling across the three dimensions of task, workspace, and harness, allowing the number of environments to grow naturally in a combined manner; constructing a unified reward system that unifies and internalizes execution-based validation, rubric evaluation of text and post-rendered visual output, and agency checks, eliminating inconsistencies in task-specific validators; and building an online data equalizer that reconstructs a balanced distribution for each batch across task, difficulty, workspace, and harness, reducing gradient variance and supporting stable and continuous expansion of training computing power.
• Self-evolutionary feedback loopThe model advances optimization through an adaptive mechanism of execution-feedback-iteration, maintaining a highly coherent systematic strategy through hundreds to thousands of interactions. It relies on feedback loops such as simulation, synthesis, and layout to achieve algorithm-level data path reconstruction rather than shallow syntax fine-tuning.
• Dynamic workflow orchestrationDrive task planning through programming, solidify orchestration logic into reproducible programs, support flexible switching from single-link coherent development to large-scale parallel exploration, and transform the serial advancement of complex tasks into an automated closed loop that can be delivered at scale within a single dialogue.

How to use Qwen 3.8-Max

• Qianwen AI Platform CallRegister an account to obtain an API Key, and directly call the model through an interface compatible with the OpenAI or Anthropic protocols.
• Adjusting the depth of reasoning:pass reasoning_effort Parameter settings xhigh(Default, deep analysis)medium(Balance) or low(Efficient reasoning), control costs and performance as needed.
• Integrated intelligent agent frameworkAfter configuring the API Key and Base URL, you can access mainstream programming assistants and agent frameworks such as Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw.
• Open source weight deploymentStarting next week, we will download model weights from Hugging Face or ModelScope for local deployment and secondary development.

The core advantages of Qwen 3.8-Max

• Ultra-long-range autonomous executionIt can focus continuously for several days to dozens of days, and evolves itself through a closed loop of execution-feedback-iteration, without the need for human intervention throughout the process.
• Real-world RL trainingIt combines the number of extended environments and training computing power, and shows balanced and continuous improvement on harnesses such as QwenWork, Claude Code, Codex, OpenClaw, and Hermes.
• End-to-end delivery capabilityBy solidifying orchestration logic into reproducible procedures through dynamic workflows, complex tasks that originally required weeks of sequential execution can be compressed into an automated closed loop that can be delivered at scale within a single conversation.

Qwen 3.8-Max's project address

• Project official website：https://qwen.ai/blog?id=qwen3.8