Qwen 3.8-Max, a language model with 2.4 trillion parameters, was announced by Qwen Studio on August 3, 2026. The model will be made available through QwenCloud, and its weights will be released as open source next week. The model builds on the Qwen 3.5 architecture and marks the first release of a Qwen-Max-class model with open weights. In a 10+ day autonomous coding run, Qwen 3.8-Max built a self-evolving harness for the open-source project oh-my-cli. The system generated 265 commits, 127 pull requests and 151 issues from July 30, 2026. It used a loop that moved tasks through a state machine, executed tests, and incorporated feedback for continuous improvement. The model reproduced a research paper titled "Unified Data Selection for LLM Reasoning" and then improved the methodology. After five days of continuous operation it wrote about 7,600 lines of code and performed 33 training cycles. The original reproduction achieved a 49.58% success rate on the AIME24 math benchmark. Through four iterative improvement rounds the model reached a 52.29% success rate, a gain of 2.71 points over the baseline. Qwen 3.8-Max entered the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba Cloud’s Tianchi platform. Working within a 24‑hour limit, the model fine‑tuned several language and vision models and submitted 45 times. Its final accuracy was 0.853, surpassing 458 of the 526 human teams. The model was tested on a range of professional tasks. It located 1,284 relevant clauses in a large legal corpus within an hour, created a high‑fidelity prototype for a banking app in one step, generated a 26‑dish restaurant menu with cost annotations, reconstructed a seismic model of a 30‑story building from drawings, produced a 3D patient education tool from a paper form, and generated tactical reports for a sports team from thousands of game logs. In an autonomous silicon design run, Qwen 3.8-Max optimized a GCD/RSA hardware accelerator. Starting with 8,298 gates, the model reduced the gate count to 678, a reduction of 92%. The final silicon layout measured 46×46 µm² and met timing requirements at 500 MHz. On the E‑commerce Bench simulation, the model managed multiple online stores for a full year, ending with a total balance of ¥416,252, a 4.16‑fold return, and a net profit of over ¥100,000 during a major promotion period, outperforming the second‑place model by 38%. Vision capabilities were demonstrated through document summarization of more than 200‑page PDFs, video memory construction for 100‑hour streams, and autonomous creation of interactive applications from screenshots or natural language. The model continuously monitored its own outputs and corrected spatial or design errors during execution. A hybrid agent benchmark, RecreationBench, evaluated the ability to rebuild running applications from black‑box observation. Qwen 3.8-Max achieved a score of 51.7, indicating frontier‑level performance in iterative coding and interface feedback loops. The model is accessible via QwenCloud using standard OpenAI‑compatible APIs. Users can adjust reasoning effort through xhigh, medium, or low settings, and preserve thinking is enabled by default. Integration is supported with Claude Code, Codex, Qoder, Qwen Code, and OpenClaw through documented APIs. Qwen 3.8-Max represents the latest advancement in the Qwen family, with open weights scheduled for release next week. Community testing and feedback are anticipated.