Qwen 3.8-Max – The Flagship Large Model Launched by the Qwen Team at Alibaba Cloud

1. What is Qwen 3.8-Max

Qwen 3.8-Max is a flagship large language model developed by the Qwen team at Alibaba Cloud, with a total parameter count of 24 trillion and 95 billion activated parameters. It is built upon the Qwen 3.5 architecture and has been extended in scale. As the first Max-level model in the Qwen series to be open-sourced, it achieves comprehensive upgrades in programming, office work, scientific research, and long-term autonomous tasks. The model supports API calls through the Qwen AI platform, is compatible with OpenAI and Anthropic protocols, and can be seamlessly integrated into mainstream agent frameworks such as Claude Code, Codex, and Qoder, demonstrating end-to-end task delivery capabilities ranging from single-turn conversations to multi-day autonomous programming.

Image source: Official article
Image source: official article

Technical positioning and domain: Belongs to the field of large-scale language models, positioned as a foundational model for general-purpose agents, with a particular emphasis on enhancing the ability to execute long-term autonomous tasks. Unlike traditional conversational models, Qwen 3.8-Max compresses complex workflows such as programming, scientific research, and office work into an automated closed-loop within a single conversation, serving as the core reasoning engine within agent frameworks.

Development background: Developed by the Qwen team at Alibaba Cloud, based on the Qwen 3.5 architecture with expanded scale. The motivation for its development stems from the capability gap in large models for autonomous execution and self-evolution over multiple days in real-world tasks, aiming to break through the current limitations of models that rely on human intervention in long-term tasks.

Core value: Addresses the challenge of requiring long-term autonomous execution for complex tasks, enabling a leap from single-turn instructions to multi-day autonomous programming. Through a real-world reinforcement learning system, the model maintains a high degree of coherence across hundreds to thousands of interactions, capable of autonomously completing professional tasks such as paper reproduction, chip design, and quantitative strategy development, significantly reducing human effort.

Technical features: Utilizes a real-world reinforcement learning system, enhancing general work capabilities by jointly expanding the number of environments and training computing power; possesses a self-evolving feedback loop, allowing for self-optimization through execution, feedback, and iteration; supports dynamic workflow orchestration, solidifying orchestration logic into reproducible programs, enabling flexible switching from single-path development to large-scale parallel exploration.

2. Key Features

• Long-term Autonomous Programming: Supports autonomous programming for up to ten days, enabling the model to build self-evolving harnesses and independently complete tasks such as requirement gathering, issue distribution, code generation, and self-repair. Achieved a score of 73.5 on the FrontierSWE benchmark, capable of iterating code independently for several consecutive days without human intervention in real project environments.

• Paper Replication and Improvement: After independently replicating the experiments in a paper, the model tested 18 improvement ideas through four rounds of self-evolution, surpassing the original method by 2.7 points on AIME24. With a PaperBench score of 93.0, it demonstrates a complete research capability, from understanding the paper to experimental replication and innovative enhancement.

• Competition Performance: Independently built and iterated on a solution within 24 hours, increasing accuracy from 0.60 to 0.853 and defeating 87% of human competitors. This feature highlights the model's rapid learning and strategic adjustment capabilities under high-pressure, time-constrained scenarios.

• Office Efficiency Enhancement: Covers hundreds of high-value professional scenarios, such as a lawyer marking 1,284 clauses within one hour or a designer creating an 8-page interactive prototype in a single attempt. Through dynamic workflow orchestration, complex tasks that originally required weeks of sequential processing are compressed into an automated closed-loop within a single conversation.

• Quantitative Strategy Development: Starting from a single-sentence description, the model coordinates approximately 330 sub-agents to complete around 6,000 backtests and deliver an end-to-end ETF rotation strategy. This showcases the model's multi-agent collaboration and large-scale parallel computing capabilities in the field of financial quantification.

• Autonomous Chip Design: Through approximately 500 rounds of interaction, the model optimized the gate count of a cryptographic accelerator netlist from 8,298 to 678, reducing the area by 81% and achieving timing closure. This demonstrates the model's deep reasoning and continuous optimization capabilities in the highly specialized domain of hardware design.

3. How to Use

1. Qwen AI Platform Invocation: Register for a Qwen AI Platform account and obtain an API Key. The model is compatible with the OpenAI and Anthropic protocols and can be invoked using standard HTTP requests. Example command: curl （link to be updated after official release） -H "Authorization: Bearer YOUR_API_KEY" -d '{"model":"qwen3.8-max","messages":[{"role":"user","content":"Hello"}]}'. It is recommended to first test the connection with small-scale requests.

2. Adjusting Reasoning Depth: Control the model's analytical depth using the reasoning_effort parameter. Available options include xhigh (default, deep analysis, suitable for complex tasks), medium (balanced, suitable for regular tasks), and low (efficient reasoning, suitable for simple Q&A). Adjust as needed to balance cost and performance; for long-range tasks, it is recommended to use xhigh.

3. Integration with Agent Frameworks: After configuring the API Key and Base URL, you can integrate with mainstream programming assistants and agent frameworks such as Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw. Specific steps: In the framework configuration file, set api_key and base_url to the values provided by the Qwen AI Platform, then start the framework to invoke Qwen 3.8-Max as the inference backend.

4. Open-Source Weight Deployment: Model weights are planned to be available for download on the Hugging Face and ModelScope platforms next week. At that time, you can use the transformers library or vLLM and other inference frameworks to load the weights for local deployment. Recommended hardware configuration: At least 4 NVIDIA A100 80GB GPUs or equivalent domestic acceleration cards, supporting FP16 inference. After downloading, run from transformers import AutoModelForCausalLM; model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-Max") to load the model.

5. Best Practices: For ultra-long-range tasks (such as multi-day programming), it is recommended to enable resume-from-breakpoint mechanisms and regularly save intermediate states; when integrating with agent frameworks, ensure a stable network environment to avoid API call timeouts; for first-time use, start with the medium reasoning depth and gradually adjust to xhigh to evaluate costs.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: If the task involves long-term autonomous execution (such as multi-day programming, chip design, or quantitative strategy development), Qwen 3.8-Max is currently the only publicly available model that supports hundreds of rounds of interaction without requiring human intervention. Its self-evolution loop and dynamic workflow orchestration capabilities provide significant advantages in this scenario. GPT 5.6 Sol performs slightly better on the terminal coding benchmark and has a more mature ecosystem, making it suitable for developers who need to quickly integrate it into existing toolchains. For multimodal reasoning, the scores are close, but Qwen leads by 16.5 points on BabyVision (infant visual understanding), indicating its potential in fine-grained visual tasks.

If the team prioritizes open-source controllability, Qwen 3.8-Max is set to open-source its weights, while GPT 5.6 Sol and Claude 4 Opus are both closed-source. Although Gemini 2.5 Pro can be accessed via API, it is not open-sourced. Therefore, Qwen 3.8-Max is a more suitable choice for enterprises with high requirements for data privacy and customization. In general office work and paper reproduction scenarios, Qwen 3.8-Max leads on both CoWorkBench and JobBench, making it ideal for research institutions and knowledge-intensive industries.

6. Editor's Summary

Qwen 3.8-Max has taken a differentiated path in technological innovation. Its real-world reinforcement learning system does not simply scale up training data, but instead decouples three dimensions—tasks, workspaces, and Harness—to allow the number of environments to grow naturally through combinations, while also building a unified reward system to eliminate inconsistencies from specialized validators. This design enables the model to maintain strategy consistency across hundreds to thousands of interactions, and its practicality has been validated in scenarios such as chip design (reducing area by 81% over 500 optimization rounds) and quantitative strategies (orchestrating 330 sub-agents). Compared to GPT 5.6 Sol's marginal advantage in terminal encoding, Qwen 3.8-Max demonstrates more breakthrough performance in long-horizon autonomous tasks and scientific research reproduction. The scores of 93.0 on PaperBench and 2.7 points above the original text on AIME24 indicate that it has already achieved independent research capabilities.

In terms of practical value, the model covers multiple high-value domains, including programming, office work, scientific research, finance, and hardware design, with specific data supporting each scenario rather than just conceptual demonstrations. Particularly notable are metrics such as "identifying 1,284 clauses within one hour" and "generating an 8-page interactive prototype in a single attempt," which directly align with enterprise needs for cost reduction and efficiency improvement. The target users include AI engineers who require automation of complex workflows, researchers needing paper reproduction and experimental acceleration, and enterprises aiming to integrate AI into core business processes. The model is set to open-source its weights, which will further lower the barrier to entry and promote community ecosystem development. In the future, as the open-source community contributes plugins and optimization solutions, Qwen 3.8-Max is expected to continue evolving in long-horizon task reliability, multimodal integration depth, and industry-specific fine-tuning, becoming a key choice for foundational agent models.

7. Application Scenarios

• Legal Compliance Review: Read through hundreds of documents in a single session and flag over a thousand relevant clauses within an hour, replacing approximately a week's worth of work for a legal team. The model automatically generates review reports through long-range context understanding and precise clause matching, supporting cross-border regulatory comparisons.

• UI/UX Prototyping: Generate an 8-page high-fidelity interactive prototype for a digital banking App in one go, eliminating the need for manual revisions. After understanding product requirements, the model directly outputs HTML/CSS/JS code, supporting real-time preview and interactive logic definition, reducing the design cycle from days to hours.

• Catering Menu Development: Based on hundreds of ingredient data sources, produce a complete menu of 26 dishes in one go, keeping the ingredient cost ratio at 33.8%. The model comprehensively considers nutritional balance, cost control, and flavor diversity to generate a complete solution that includes cooking steps and supply chain recommendations.

• Structural Engineering Modeling: Reconstruct a 30-story office building's seismic structure model in the browser based solely on a single blueprint, supporting real-time hover viewing of key metrics. After parsing the blueprint, the model automatically generates a 3D model, annotating beam and column nodes and load parameters to assist engineers in seismic analysis.

• Rehabilitation Medicine Visualization: Convert 2D paper-based assessment forms into freely rotatable, layered 3D interactive demonstrations to help patients understand their rehabilitation journey. The model generates personalized rehabilitation animations using a medical knowledge graph, supporting joint angle annotations and progress tracking.

8. FAQ

Q: How does Qwen 3.8-Max compare to GPT 5.6 Sol?
A: Both models excel in different dimensions. Qwen 3.8-Max outperforms in long-horizon autonomous tasks (such as chip design, quantitative strategies), paper reproduction (PaperBench 93.0 vs 90.5), and general office tasks (CoWorkBench 74.8 vs 71.5); GPT 5.6 Sol performs slightly better in terminal coding (Terminal Bench 88.8 vs 86.6) and multimodal reasoning (MMMU-Pro 83.0 vs 82.3). The choice depends on the specific task type.

Q: How to obtain an API Key for Qwen 3.8-Max?
A: Register an account on the Qwen AI platform (qwen.ai) and create an API Key in the console. The model is compatible with OpenAI and Anthropic protocols, so existing client libraries can be used directly by modifying the base_url and api_key.

Q: When will the open-source weights be released? What hardware is required?
A: The model weights are planned to be open-sourced next week on the Hugging Face and ModelScope platforms. The recommended hardware configuration is at least 4 NVIDIA A100 80GB GPUs or equivalent computing power, supporting FP16 inference. Due to the total parameter count of 24 trillion and active parameters of 95 billion, consumer-grade GPUs cannot run the model.

Q: Does Qwen 3.8-Max support Chinese?
A: The model is developed based on the Qwen series, which itself supports both Chinese and English. However, the official benchmark tests mainly focus on English tasks, and the specific performance in Chinese scenarios needs to be validated by the community after the open-source release. It is recommended to conduct small-scale testing for Chinese tasks first.

Q: Can it be used for commercial purposes?
A: The model will be released under an open-source license, with the specific license type to be officially announced. Historical Qwen models have mostly used Apache 2.0 or similar permissive licenses, and it is expected that Qwen 3.8-Max will also support commercial use. API calls are subject to the service terms of the Qwen AI platform.

9. Project Links

• Official Project Announcement: https://qwen.ai/blog?id=qwen3.8
• Qwen Team Homepage: https://qwen.ai
• Hugging Face Qwen Organization: https://huggingface.co/Qwen
• ModelScope Qwen Organization: https://modelscope.cn/organization/qwen