{"slug": "alibaba-launches-qwen3-8-max-with-open-weights-scheduled-next-week", "title": "Alibaba launches Qwen3.8-Max, with open weights scheduled next week", "summary": "Alibaba released Qwen3.8-Max on August 2nd, a mixture-of-experts model with 2.4 trillion total parameters and 95 billion active parameters, available immediately through QwenCloud with open weights promised the following week. The model, developed under Alibaba Cloud CTO Jingren Zhou, is priced at $2 per million input tokens and $6 per million output tokens, with a 1 million token context window. Alibaba positions it for coding, research, and multi-day tasks, and it demonstrated autonomous coding by building a project with 265 commits and 127 pull requests as of July 30th.", "body_md": "Alibaba released [Qwen3.8-Max](https://qwen.ai/blog?id=qwen3.8) on August 2nd, giving developers immediate access through QwenCloud and promising to publish the model's weights the following week. The release marks Alibaba's first plan to open a Max-class Qwen model, extending its open-weight strategy to the most computationally ambitious tier in the family. ([qwen.ai](https://qwen.ai/blog?id=qwen3.8))\n\nQwen3.8-Max is a mixture-of-experts model with 2.4 trillion total parameters and 95 billion active parameters, according to Qwen. Alibaba is positioning it for coding, research, professional work and tasks that continue across multiple days, rather than short question-and-answer sessions.\n\nThe model work sits inside Alibaba Cloud under [chief technology officer](/article/rob-witoff-joins-coinbase-as-cto) [Jingren Zhou](https://www.columbia.edu/content/events/lecture-series-ai-jingren-zhou-cto-alibaba-cloud), who leads the development of Qwen and Alibaba's Wan video models. Zhou previously worked on database and big-data systems at Microsoft and holds a computer science doctorate from Columbia University. His remit connects Qwen's model development directly to the cloud infrastructure Alibaba wants developers to pay to use. ([columbia.edu](https://www.columbia.edu/content/events/lecture-series-ai-jingren-zhou-cto-alibaba-cloud))\n\n### Hosted first, weights later\n\nQwen3.8-Max is already available through [QwenCloud](https://www.qwencloud.com/models/qwen3.8-max), including an OpenAI-compatible API. QwenCloud lists a context window of 1 million tokens, maximum output of roughly 131,000 tokens and support for text, image and video inputs. The listed price is $2 per million input tokens and $6 per million output tokens, with lower rates for cached input. ([docs.qwencloud.com](https://docs.qwencloud.com/developer-guides/getting-started/introduction))\n\nThat sequencing gives Alibaba a head start on cloud usage before developers can download and run the model elsewhere. The hosted endpoint also provides the simplest route for teams that lack the infrastructure needed to serve a model with 95 billion active parameters.\n\nAlibaba has not attached a valuation or venture round to Qwen because the model family is funded inside the group. In February 2025, Alibaba committed at least RMB380 billion, then about $53 billion, to cloud and AI infrastructure over three years. Chief executive [Eddie Wu](https://home.alibabagroup.com/en-US/about-alibaba-leadership-1637927598568767488), an Alibaba co-founder and its first technology director, has made cloud computing the commercial center of that spending plan. ([alibabagroup.com](https://www.alibabagroup.com/en-US/document-1830678592242057216))\n\nThe economics are already visible in Alibaba's reporting. In March, Alibaba said Cloud Intelligence Group revenue for the December quarter rose 36% year over year to RMB43.3 billion, while AI-related product revenue recorded triple-digit growth for a tenth consecutive quarter. Those are company-reported figures, and Alibaba does not separate Qwen model revenue from its broader cloud results. ([home.alibabagroup.com](https://home.alibabagroup.com/en-US/document-1971445322303406080))\n\n### Qwen makes autonomous coding the launch case\n\nAlibaba's central claim for Qwen3.8-Max concerns sustained execution. In one demonstration, Qwen says the model spent about 16 days building and maintaining an autonomous code-agent project called `oh-my-cli`\n\n. Requirements entered through GitHub issues, agents claimed the work, and changes proceeded through testing and pull-request checks.\n\nThe source announcement counted 265 commits, 127 pull requests and 151 issues as of July 30th. By August 3rd, the public repository showed 424 commits. Its code and operating rules are inspectable, including an \"autonomy contract\" that describes how user reports, community findings and reproducible model discoveries become normalized issues for execution. The repository uses Node.js and TypeScript and carries an Apache 2.0 license. ([github.com](https://github.com/qwen-code-dev-bot/oh-my-cli))\n\nQwen also says the model reproduced the experiments from a recent data-selection research paper and then attempted to improve on the published method. According to the launch post, the run lasted about 125 hours, generated roughly 7,600 lines of code, took more than 1,100 actions and completed 33 GPU training runs without human help.\n\nThose demonstrations are stronger evidence than a static coding benchmark because they leave behind code and project history. They still originate with Qwen, and the launch case rests heavily on Alibaba's own evaluations and selected project traces. The performance comparisons in the announcement should be read as company results until outside evaluators test the hosted model and, later, the downloadable weights.\n\nThe open-weight commitment is the strategic hinge. Alibaba has used downloadable Qwen models to gain distribution outside its own products, while QwenCloud converts a portion of that developer attention into paid inference. Alibaba said in March that the Qwen family had exceeded 1 billion cumulative downloads on Hugging Face as of January 21st. ([home.alibabagroup.com](https://home.alibabagroup.com/en-US/document-1971445322303406080))\n\nQwen3.8-Max brings that playbook to Alibaba's flagship tier. The API is the product available today. The promised weights will determine how far developers can inspect, modify and independently deploy the model, and whether Alibaba has delivered an open alternative at the same tier it sells through its cloud.", "url": "https://wpnews.pro/news/alibaba-launches-qwen3-8-max-with-open-weights-scheduled-next-week", "canonical_source": "https://runtimewire.com/article/alibaba-launches-qwen3-8-max-open-weights-next-week", "published_at": "2026-08-03 04:03:19+00:00", "updated_at": "2026-08-03 04:27:04.430372+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-infrastructure"], "entities": ["Alibaba", "Qwen3.8-Max", "QwenCloud", "Jingren Zhou", "Eddie Wu", "Alibaba Cloud", "oh-my-cli"], "alternates": {"html": "https://wpnews.pro/news/alibaba-launches-qwen3-8-max-with-open-weights-scheduled-next-week", "markdown": "https://wpnews.pro/news/alibaba-launches-qwen3-8-max-with-open-weights-scheduled-next-week.md", "text": "https://wpnews.pro/news/alibaba-launches-qwen3-8-max-with-open-weights-scheduled-next-week.txt", "jsonld": "https://wpnews.pro/news/alibaba-launches-qwen3-8-max-with-open-weights-scheduled-next-week.jsonld"}}