Qwen3.8-Max release notes / what's new

Qwen has announced the release of Qwen3.8-Max, the most capable model in the Qwen family to date. Scaling to 2.4 trillion parameters (with 95B active), the model is designed to handle complex, open-ended goals end-to-end, specifically targeting improvements in autonomous coding, professional research, and long-horizon task execution.

Autonomous Coding and Self-Evolution

Qwen3.8-Max transitions from simple function generation to autonomous project management. The model demonstrates the ability to self-evolve through feedback loops, as evidenced by three primary test cases:

  • Self-Evolving Harness: In a 10+ day autonomous run, the model built the oh-my-cli project from scratch. It implemented a loop involving an issue state machine, dispatcher, and watchdog to convert community feedback into executable work. After 16 days, the repository accumulated 265 commits, 127 PRs, and 151 issues.
  • Research Reproduction and Improvement: The model reproduced the "Unified Data Selection for LLM Reasoning" paper from scratch (writing 7,600 lines of code over 125 hours) and subsequently evolved a new method that outperformed the original paper's approach, achieving a +2.7 point gain on the AIME24 benchmark.
  • Competitive Programming: In the WWW2025 Multimodal Dialogue Intent Recognition Challenge, Qwen3.8-Max competed against 526 human teams. Working autonomously within a 24-hour limit, it achieved an accuracy of 0.853, beating 87% of the human participants.

Professional Workflow Integration

To improve general working competence, Qwen scaled RL environments and compute across several harnesses, including QwenWork, Claude Code, and Codex. This was achieved through a universal reward system that unifies execution-based checking and agentic inspection.

High-Value Profession Performance

Qwen3.8-Max was stress-tested across hundreds of professions to deliver production-quality results:

  • Legal: Completed a corporate compliance review of hundreds of documents in under an hour, a task typically requiring a paralegal team a full week.
  • UI/UX Design: Produced a high-fidelity, 8-screen interactive prototype for a banking app in one shot without human revision.
  • Engineering: Reconstructed a seismic structural model of a 30-story office tower from drawings, providing real-time inspection of base shear and inter-story drift.
  • Medical: Converted a 2D paper assessment into a 3D interactive anatomical demo, reducing lead times from weeks to minutes.
  • Finance: Developed an end-to-end ETF-rotation strategy, autonomously pruning redundant factors to avoid overfitting and executing massively parallel factor mining via 330 sub-agents.

Long-Horizon Task Execution

Qwen3.8-Max utilizes an action-feedback-iteration loop to handle tasks requiring thousands of rounds of interaction.

Autonomous Chip Design

The model executed the full silicon design flow for a GCD/RSA cryptographic hardware accelerator. Without reference designs, it used an "edit-simulate-synthesize-layout" loop to reduce the design from 8,298 gates to 678 gates. The final physical layout achieved an 81% reduction in die area and successfully reached timing closure at 500 MHz.

E-Commerce Simulation

On the E-Commerce Bench (a 365-day simulation), Qwen3.8-Max managed multiple stores, handled supply chain crises, and negotiated with suppliers using game theory principles. It achieved a total balance of ¥416,252 (a 4.16x return), surpassing the second-place model (GLM 5.2) by 38%.

Multimodal Intelligence and Hybrid Agents

Qwen3.8-Max integrates visual intelligence into the entire task lifecycle, moving beyond simple image understanding to active visual feedback.

  • Visual Feedback Loops: The model can observe its own intermediate results (e.g., page layouts or animation quality) and autonomously revise its plan to correct deviations.
  • Hybrid Agent Capability: By pairing coding with GUI operation, the model can implement logic via code while verifying results against a live system. This was tested via RecreationBench, where the model rebuilt applications from black-box observation.
  • Video Memory: For videos exceeding 100 hours, the model constructs a video memory graph to track characters, events, and relationships over long durations.

Technical Specifications and Availability

  • Parameters: 2.4 Trillion (95B active).
  • Availability: Currently available via QwenCloud.
  • Open Weights: Scheduled for release next week on Hugging Face and ModelScope.
  • API Features: Supports reasoning_effort (xhigh, medium, low) to balance depth and cost, and preserve_thinking for transparency in reasoning processes.

Sources