Alibaba has launched Qwen3.8-Max, an open-weight language model with 2.4 trillion parameters designed to complete complex tasks independently over multiple days.
The system currently runs on QwenCloud, with full weights scheduled for release on Hugging Face and ModelScope next week. It supports both OpenAI’s Chat Completions format and Anthropic’s API protocol.
Autonomous coding and research
The development team demonstrated the model’s capabilities through three specific case studies involving zero human intervention.
In the first instance, Qwen3.8-Max spent 16 days building the command-line tool oh-my-cli. The model accepted user requests, converted them into GitHub issues, assigned tasks to itself, wrote code, ran tests, and iterated on improvements. By July 30, 2026, the project had accumulated 265 commits, 127 pull requests, and 151 issues.
The second test involved reproducing and improving on the results of the research paper “Unified Data Selection for LLM Reasoning”. The model received the paper but no starter code. Over roughly five days and approximately 125 hours of compute time, it generated 7,600 lines of code and executed 33 GPU training jobs. It successfully reproduced all six main results from the paper. Following this, it tested 18 of its own ideas across four rounds, beating the original method on the AIME24 math benchmark by 2.7 points.
The third case covered the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba’s Tianchi platform, where 526 human teams competed. Within 24 hours, the model fine-tuned several Chinese language models alongside Qwen2.5-VL-7B for product screenshots and combined them into a voting system. Across 45 submissions, accuracy rose from 0.60 to 0.853. This performance placed Qwen3.8-Max ahead of 458 of the human teams.
Chip design and business simulation
Two further case studies examined tasks requiring hundreds of interaction rounds.
The first required designing a cryptographic building block for encryption schemes. The key efficiency metric for such a circuit is the number of logic gates required, with fewer gates indicating a smaller, more efficient chip. The model began with a functional but bloated design using 8,298 gates. It reduced this to 678 gates over approximately 500 iterations.
After an automated layout pass using the open-source tool OpenROAD, the physical chip area shrank from 106×106 to 46×46 micrometers, an 81 percent reduction. The Qwen team noted the model made deep structural changes even after hundreds of iterations rather than settling for surface-level adjustments.
The second case was E-Commerce-Bench, a simulation of an entire fiscal year in online retail based on anonymized data from Taobao and Tmall. The model started with 100,000 yuan in capital and had to run multiple online stores in parallel for a full year. This involved buying products, negotiating with suppliers in natural language, adjusting prices, managing returns, and handling crises like typhoons or supply chain disruptions.
Hidden within the supplier pool were 152 scammers that the model had to identify. Qwen3.8-Max finished with a balance of 416,252 yuan, quadrupling its starting capital. This result was 38 percent more than the runner-up GLM 5.2 and more than 2.5 times what its predecessor Qwen3.7-Max managed. The model invested aggressively early in the year and recorded a net profit of over 100,000 yuan during the holiday season.
Multimodal skills and app reconstruction
For multimodal tasks, the model can process documents exceeding 200 pages and videos longer than 100 hours. The team is also introducing RecreationBench, a new benchmark requiring the model to rebuild running applications without access to source code.
The model observes the target app only through interaction, meaning clicks and keyboard input. Testing covers Ubuntu, macOS, Windows, Android, and the web. Alongside this, the team is releasing Qwen-MM-Plugins, an extension library adding image and video processing, visual tool use, and multimodal memory to existing agent systems.
In the benchmark tables the Qwen team published, the model lands near or above Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Sol across many categories. On PaperBench, Qwen3.8-Max hits 93, the highest score in the comparison. On TerminalBench 2.1, it scores 86.6, trailing GPT-5.6 Sol’s 88.8. As is typical with self-reported numbers from model makers, these results come from internal runs. Independent verification is still pending.
The Qwen team attributes the model’s ability to sustain such long-running tasks to a major expansion of training environments during reinforcement learning. Training no longer focused on single tasks alone but also covered multi-day workflows, nested directory structures instead of individual files, and a variety of agent harnesses.
The team’s 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 dipped slightly.
China’s open-model race heats up
Qwen3.8-Max’s most direct rival also comes from China. Moonshot AI released Kimi K3 with open weights on Hugging Face on July 27, a multimodal mixture-of-experts model with 2.8 trillion parameters and a one-million-token context window. Along with the weights, Moonshot also published parts of its own infrastructure, including attention kernels, an MoE communication library, and tools for running agents at scale. Independent testing tempered the company’s claims, though. K3 fell well short of top Western models in both cyber capabilities and complex math.
A parameter called
reasoning_effort
lets users choose between three levels that trade speed for thoroughness.
Qwen3.8-Max is available now through QwenCloud. The weights are set to go live on Hugging Face and ModelScope next week. The model supports both OpenAI’s Chat Completions format and Anthropic’s API protocol, so it plugs directly into Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw.
Qwen3.8-Max isn’t the only piece Alibaba has added recently. A few weeks ago, the team introduced Qwen-Image-3.0, an image generator for information-dense layouts that handles inputs up to 4,500 tokens and renders readable text as small as ten pixels. At the other end of the scale, Alibaba continues to push small open models like Qwen3.6-35B-A3B, which has 35 billion total parameters but only activates three billion at a time.
What it means
Developers can now run this model on their own infrastructure to handle workflows that span days without constant supervision. The tool allows teams to automate repetitive coding cycles, reproduce research findings, and manage complex business logic where a single prompt is insufficient.
