Qwen 3.8-Max – The flagship large model launched by Alibaba Cloud’s Qwen team Operational

Qwen 3.8-Max is the flagship large model launched by Alibaba Cloud’s Qwen team, with a total of 2.4 trillion parameters and an architecture expanded from Qwen 3.5. As the most powerful model in the Qwen family to date, it is the first Max-level model set to release its weights as open source. The model delivers comprehensive upgrades for coding, office work, research, and long-horizon tasks. It can be called through the Qwen AI platform API, is compatible with the OpenAI and Anthropic protocols, and integrates seamlessly with mainstream agent frameworks such as Claude Code, Codex, and Qoder.

Long-horizon autonomous coding: Supports autonomous coding over ten-day periods, building a self-evolving harness to complete requirements gathering, issue assignment, code generation, and self-repair.
Paper reproduction and improvement: Independently reproduces paper experiments, then tests 18 improvement ideas through four self-evolution cycles, surpassing the original method by 2.7 points on AIME24.
Competition performance: Independently builds and iterates on solutions within 24 hours, raising accuracy from 0.60 to 0.853 and defeating 87% of human participant teams.
Workplace productivity: Covers hundreds of high-value professional scenarios, such as a lawyer identifying 1,284 clauses in one hour and a designer producing an 8-page interactive prototype in one pass.
Quantitative strategy development: Starting from a single-sentence description, orchestrates approximately 330 sub-agents, completes approximately 6,000 backtests, and delivers an end-to-end ETF rotation strategy.
Autonomous chip design: After approximately 500 interaction rounds, reduces the gate count of a cryptographic accelerator netlist from 8,298 to 678, cuts area by 81%, and achieves timing closure.

Architecture expansion: Expanded from the Qwen 3.5 architecture, with a total of 2.4 trillion parameters and 95B activated parameters. It achieves a major increase in scale while retaining architectural continuity.
Real-world reinforcement learning system: Improves general work capabilities by jointly scaling the number of RL environments and training compute. It addresses three coupled challenges: decoupling tasks, workspaces, and harnesses so the number of environments can grow naturally through combination; building a unified reward system that internalizes execution-based verification, rubric evaluation of text and rendered visual outputs, and agentic checks, eliminating inconsistencies among task-specific validators; and building an online data balancer that reconstructs balanced distributions for each batch across tasks, difficulty, workspaces, and harnesses, reducing gradient variance and supporting the stable, continuous expansion of training compute.
Self-evolving feedback loop: The model advances optimization through an adaptive execution-feedback-iteration mechanism, maintaining highly coherent and systematic strategies over hundreds to thousands of interaction rounds. It relies on feedback loops involving simulation, synthesis, placement, and other processes to achieve algorithm-level data-path restructuring rather than superficial syntax tuning.
Dynamic workflow orchestration: Uses programming to drive task planning and solidifies orchestration logic into reproducible programs. It supports flexible switching from coherent single-chain R&D to large-scale parallel exploration, transforming the serial progress of complex tasks into an automated closed loop that can be delivered at scale within a single conversation.

Call through the Qwen AI platform: Register an account and obtain an API Key, then call the model directly through an interface compatible with the OpenAI or Anthropic protocols.
Adjust reasoning depth: Set the reasoning_effort parameter to xhigh (default, deep analysis), medium (balanced), or low (efficient reasoning) to control cost and performance as needed.
Integrate agent frameworks: After configuring the API Key and Base URL, connect to mainstream coding assistants and agent frameworks such as Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw.
Deploy open-source weights: Starting next week, download the model weights from Hugging Face or ModelScope for local deployment and further development.

Ultra-long autonomous execution: Maintains focused execution continuously for days to tens of days and self-evolves through an execution-feedback-iteration loop without human intervention.
Real-world RL training: Jointly scales the number of environments and training compute, delivering balanced and steadily improving performance on harnesses including QwenWork, Claude Code, Codex, OpenClaw, and Hermes.
End-to-end delivery capability: Solidifies orchestration logic into reproducible programs through dynamic workflows, transforming complex tasks that originally required weeks of serial progress into an automated closed loop that can be delivered at scale within a single conversation.

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

Legal and compliance review: Reads hundreds of documents in a single pass and identifies thousands of relevant clauses within one hour, replacing approximately one week of work by a legal team.
UI/UX prototyping: Generates an 8-page high-fidelity interactive prototype for a digital banking app in one pass, with no manual rework required.
Restaurant menu development: Produces a complete menu of 26 dishes in one pass based on more than 100 ingredient documents, while controlling the ingredient cost ratio at 33.8%.
Structural engineering modeling: Reconstructs the seismic structural model of a 30-story office building in a browser using only one drawing, with real-time hover access to key metrics.
Rehabilitation medicine visualization: Converts a 2D paper assessment form into a freely rotatable 3D interactive presentation with layer-by-layer reveal, helping patients understand their rehabilitation pathway.