Alibaba Group's Qwen team officially released Qwen 3.8-Max on August 3rd. This model features a total of 2.4 trillion parameters with 95 billion activated parameters, making it the largest and most powerful model in the Qwen family to date. For the first time, the Max-level model is being open-sourced, with its weights scheduled for release next week on Hugging Face and ModelScope.
In terms of pricing, Qwen 3.8-Max is available via API through the Qwen AI platform on Alibaba Cloud. The cost is set at $2.0 per million input tokens and $6.0 per million output tokens, with a reduced rate of $0.25 per million tokens for implicit cache hits.
Benchmark tests reveal that Qwen 3.8-Max matches or surpasses Anthropic's Claude Fable 5 on several key metrics for coding agents and general-purpose agents. For instance, it scored 93.0 on the PaperBench, exceeding Fable 5's 88.8. On CoWorkBench, it achieved 74.8, close to Fable 5's 75.9, and on WideSearch, it scored 81.9, essentially tying with Fable 5's 81.2.
Performance: On Par with Claude Fable 5, Exceeding in Several Areas
Based on comprehensive benchmark data released by the Qwen team, Qwen 3.8-Max matches or exceeds Anthropic Claude Fable 5 on several core indicators. On PaperBench, which tests research paper capabilities, Qwen 3.8-Max scored 93.0, surpassing Fable 5's 88.8 and GPT-5.6 Sol's 90.5. On IFBench, the model scored 82.8, significantly outperforming Fable 5 at 63.5 and GPT-5.6 Sol at 72.7. In HealthBench, Qwen 3.8-Max achieved a score of 60.2, exceeding GPT-5.6 Sol's 55.3. For general collaboration on CoWorkBench, it scored 74.8, a very small margin behind Fable 5's 75.9. On JobBench, which measures AI job performance, Qwen 3.8-Max scored 53.4, close to Fable 5's 57.4.
The model also performs strongly in multimodal tasks. On OSWorld-Verified, it scored 86.1, exceeding Fable 5's 85.0. On AndroidWorld, it scored 85.3 compared to Fable 5's 88.8. On MLVU, which tests long video understanding, it achieved 90.8, a metric for which Fable 5 has no available data. Notably, the Qwen team's annotations note that some of Fable 5's results may include fallback mechanisms.
Coding Capabilities: From Empty Folder to Production-Grade Delivery
The Qwen team tested Qwen 3.8-Max's autonomous coding capabilities using three real-world scenarios with no human intervention. In the first case, which involved a ten-day automated programming project, the model created an "oh-my-cli" project from scratch. Operating autonomously for approximately 16 days, it completed 265 commits, 127 pull requests, and 151 issues. It integrated user feedback, community practices, and self-testing results into a continuous engineering loop, achieving a complete closed cycle of task assignment, code generation, and test verification.
In the second case, the model worked independently for about 125 hours to replicate and surpass a research paper. It wrote approximately 7,600 lines of code, executed over 1,100 steps, and ran 33 rounds of GPU training to fully reproduce the six main conclusions of the paper "Unified Data Selection for LLM Reasoning." Entering a "self-evolution" phase, it proposed and tested 18 improvement schemes, ultimately boosting performance on the competitive math benchmark AIME24 by 2.7 points over the paper's original method.
For the third case, the model outperformed 87% of human teams within 24 hours. Competing in the WWW2025 multimodal dialogue intent recognition challenge against 526 human teams, it iterated through 45 submissions, improving its accuracy from 0.60 to 0.853 and defeating 458 teams.
Office and Long-Horizon Tasks: Production-Grade Validation Across Hundreds of Scenarios
The Qwen team validated Qwen 3.8-Max's real-world delivery capabilities across hundreds of high-value professional scenarios. A corporate compliance lawyer task involved reading through hundreds of documents in a single pass, identifying 1,284 relevant clauses within one hour—a workload that would typically require a legal team about a week. A UI/UX designer task saw the model generate a high-fidelity, interactive prototype with eight pages without any manual revisions. For a structural engineer scenario, the model reconstructed a 30-story office building's seismic structural model in a browser based on a single blueprint, a process that traditionally takes over a week. In a quantitative research task, starting from a single sentence description, the model dispatched approximately 330 sub-agents and completed about 6,000 backtests, compressing a process that usually takes weeks into a single conversation.
On the E-Commerce Bench, a 365-day e-commerce business simulation, Qwen 3.8-Max generated ¥416,252 in profit, a 4.16x return. This result outperformed the second-place model, GLM 5.2, by 38% and represented a 152% improvement over the previous generation, Qwen 3.7-Max.
Chip Design: 500 Interaction Rounds, 81% Area Reduction
Qwen 3.8-Max demonstrated its long-horizon autonomous planning abilities in a chip design optimization task. The objective was to minimize the logic gate count of a G CD/RSA cryptographic hardware accelerator. Without any reference design or human intervention, the model engaged in approximately 500 rounds of interaction and 71 evaluations across 13 key milestones. It optimized the initial design from 8,298 gates down to 678 gates, achieving the best performance among all evaluated models. In terms of physical implementation, the chip area shrank from 106×106 µm² to 46×46 µm², and the routing length decreased from 33,369 µm to 4,187 µm. The design achieved timing closure at a frequency of 500 MHz, resulting in an overall chip area reduction of 81%.
Multimodal Capabilities: Vision Driving the Full Execution Process
The visual capabilities of Qwen 3.8-Max extend beyond simply "understanding images" to serving as a continuous feedback loop throughout task execution. During a task, the model continuously observes intermediate outputs—checking page layouts, object orientations, and animation effects—and autonomously identifies and corrects deviations when it finds problems. The Qwen team defines this capability as a "native visual feedback loop." In benchmarks, Qwen 3.8-Max scored 86.1 on OSWorld-Verified, surpassing Fable 5's 85.0. On AndroidWorld, it scored 85.3, close to Fable 5's 88.8, and on ParametricCAD Bench, it achieved 91.5, exceeding Fable 5's 87.5. To facilitate developer integration, the Qwen team has also launched Qwen-MM-Plugins, which provide extended support for image and video processing, multimodal memory, and visual tool calls across different agent frameworks.
Open Access: Supports Mainstream Frameworks like Claude Code and Codex
Qwen 3.8-Max is now available via API on the Qwen AI platform, supporting the OpenAI and Anthropic compatible protocols. This allows direct integration with mainstream development tools such as Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw. The API supports a "reasoning_effort" parameter to adjust the depth of inference: "xhigh" (default, suitable for complex tasks), "medium" (balancing accuracy and speed), and "low" (optimizing for speed and cost). The model's weights are scheduled for open-source release next week on Hugging Face and ModelScope, with the Qwen 3.8-27B model also being released concurrently.
Comments