Alibaba’s Qwen team has introduced Qwen3.8-Max, a new flagship language model aimed at a harder class of AI work: tasks that require sustained execution, iteration, and planning over long periods.
The model has 2.4 trillion total parameters, with 95 billion active per query. It builds on the Qwen3.5 architecture and is positioned less as a chatbot for isolated answers and more as an agentic system for complex workflows.
A model built for work that lasts days
Qwen3.8-Max is designed to operate across extended tasks, including software development, research reproduction, chip design, and business simulation. Alibaba first announced a preview in mid-July through Alibaba's Token Plan, Qoder, and QoderWork at ten percent of the standard price.
At that stage, the team said the model had 2.4 trillion parameters and placed it just behind Fable 5, but did not publish benchmarks. The new release is notable because Qwen3.8-Max is the first model in the Qwen-Max class whose weights will be made publicly available.
The model is available now through QwenCloud. Its weights are set to go live on Hugging Face and ModelScope next week.
Coding tests focused on autonomy
Alibaba presented three coding-focused case studies where Qwen3.8-Max worked without human help. The examples are meant to show whether the model can break down work, create artifacts, test its own output, and keep improving across many steps.
In one run, Qwen3.8-Max spent 16 days building a command-line tool called oh-my-cli. It handled incoming user requests, converted them into GitHub issues, assigned the work to itself, wrote code, ran tests, and revised the project. By July 30, 2026, the project had 265 commits, 127 pull requests, and 151 issues, according to the source account.
A second test gave the model the paper "Unified Data Selection for LLM Reasoning" but no starter code. Over roughly five days and about 125 hours of compute time, Qwen3.8-Max produced 7,600 lines of code and ran 33 GPU training jobs. It reproduced all six main results from the paper, then tried 18 of its own ideas across four rounds and beat the paper's method on the AIME24 math benchmark by 2.7 points.
The third coding case used the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba's Tianchi platform, where 526 human teams competed. In 24 hours, Qwen3.8-Max fine-tuned several Chinese language models along with Qwen2.5-VL-7B for product screenshots, then combined them into a voting system. Across 45 submissions, accuracy rose from 0.60 to 0.853, placing it ahead of 458 of the 526 human teams.
Chip design and simulated commerce test planning
Alibaba also highlighted two longer-horizon scenarios that required hundreds of interaction rounds or a full simulated operating cycle. These examples matter because they move beyond producing code into planning, optimization, and response to changing conditions.
In one case, Qwen3.8-Max worked on a cryptographic building block for encryption schemes. The task was to reduce the number of logic gates, since fewer gates can mean a smaller and more efficient chip. The model began with a working design using 8,298 gates and reduced it to 678 gates over roughly 500 iterations.
After an automated layout pass using the open-source tool OpenROAD, the physical chip area fell from 106x106 to 46x46 micrometers, an 81 percent reduction. The Qwen team says the model continued to make deep structural changes after hundreds of iterations, instead of only making surface-level adjustments.
The other case was E-Commerce-Bench, a simulation of a fiscal year in online retail using anonymized data from Taobao and Tmall. The model began with 100,000 yuan in capital and had to run multiple online stores in parallel for a full year.
That simulation required a range of business actions:
- buying products
- negotiating with suppliers in natural language
- adjusting prices
- managing returns
- responding to crises such as typhoons or supply chain disruptions
The supplier pool included 152 scammers that the model needed to identify. Qwen3.8-Max finished with a balance of 416,252 yuan, more than quadrupling the initial capital. That result was 38 percent higher than the runner-up GLM 5.2 and more than 2.5 times the result from Qwen3.7-Max. The source account says the model invested aggressively early in the year and earned a net profit of over 100,000 yuan during the holiday season.
Multimodal work and app reconstruction
Qwen3.8-Max is also presented as a multimodal system. The Qwen team says it can process documents with over 200 pages and videos longer than 100 hours.
The team is introducing RecreationBench, a benchmark where the model must rebuild running applications without source code. The model can observe the target app only through interaction, including clicks and keyboard input. Testing spans Ubuntu, macOS, Windows, Android, and the web.
Alongside the benchmark, Alibaba is releasing Qwen-MM-Plugins. The extension library adds image and video processing, visual tool use, and multimodal memory to existing agent systems.
Benchmarks, access, and the open-model race
In benchmark tables published by the Qwen team, Qwen3.8-Max lands near or above Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Sol across many categories. It scores 93 on PaperBench, the highest score in the comparison. On TerminalBench 2.1, it scores 86.6, behind GPT-5.6 Sol's 88.8.
Those results are still internal. The source notes that independent verification is pending, which is important for interpreting performance claims from any model maker.
The Qwen team attributes the model’s longer-task performance to a broader set of training environments during reinforcement learning. Instead of training only on single tasks, the process included multi-day workflows, nested directory structures rather than individual files, and multiple agent harnesses.
Alibaba says its internal score index across more than ten benchmarks rose from 0.474 to 0.725. The model performed best at around 4,000 environments, after which scores declined slightly.
Qwen3.8-Max arrives as competition among open models from China intensifies. Moonshot AI released Kimi K3 with open weights on Hugging Face on July 27. That model is described as a multimodal mixture-of-experts system with 2.8 trillion parameters and a one-million-token context window. Moonshot also published parts of its infrastructure, including attention kernels, an MoE communication library, and tools for running agents at scale. Independent testing later found K3 fell short of top Western models in cyber capabilities and complex math.
For developers, Qwen3.8-Max supports OpenAI's Chat Completions format and Anthropic's API protocol. That means it can plug into Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw. A parameter called reasoning_effort lets users choose between three levels that trade speed for thoroughness.