# Alibaba releases Qwen3.8-Max weights without clear license terms > Source: > Published: 2026-08-17 12:10:51+00:00 Alibaba released Qwen3.8-27B and opened [Qwen3.8-Max](/models/qwen/qwen3.8-max) weights on August 17, according to [CNBC](https://www.cnbc.com/2026/08/17/alibaba-meta-qwen-open-weight-ai-laptop-models.html?ref=runtimewire). [Qwen](https://qwen.ai/?ref=runtimewire), Alibaba's AI model family, now gives developers an Apache-licensed 27B model with a detailed repository and access to the flagship's learned parameters. The newly downloadable Max weights are the central development. Alibaba described Qwen3.8-Max as its most powerful model, and the August 17 release made its weights available for download and execution, according to CNBC. Qwen3.8-Max's license and official parameter specifications remain unresolved in the reviewed materials. [Qwen on X](https://x.com/Alibaba_Qwen/status/2088280182356611304?ref=runtimewire) Alibaba co-founder and CEO [Eddie Wu](https://home.alibabagroup.com/en-US/about-alibaba-leadership-1637927598568767488?ref=runtimewire) leads the group behind Qwen. Wu served as technology director when Alibaba was founded in 1999. The available sources establish the release as an Alibaba and Qwen team decision; they do not show Wu personally directing its two-model structure. Opening model weights gives developers access to the learned numerical parameters used for inference. Depending on the license and technical files, developers may be able to inspect and self-host a model. A weights release does not disclose the training data or full training process. ### The 27B repository provides the documentation Max lacks The [Qwen3.8-27B repository](https://huggingface.co/Qwen/Qwen3.8-27B?ref=runtimewire) documents a 27B multimodal model with a 262,144-token native context window that can be extended to 1 million tokens through RoPE scaling. The scaling requires configuration or serving changes and can affect performance on shorter inputs, according to the model card. Alibaba published Qwen3.8-27B under the Apache 2.0 license. Its repository provides instructions for Transformers, vLLM, SGLang and Docker. It also documents controls that let developers adjust reasoning depth or disable thinking mode. The broader [Qwen GitHub repository](https://github.com/QwenLM/Qwen3?ref=runtimewire) contains the team's public Qwen3 implementation and model resources. The Qwen3.8-27B model card lists 27,781,427,952 BF16 parameters and a total file size of 55,631,189,992 bytes. It does not specify a standard laptop configuration or measured generation speed, leaving developers to determine the hardware and quantization needed for local use. Qwen3.8-Max has a thinner public record in the materials reviewed for this article. [S5 Labs states](https://s5labs.io/resources/insights/qwen38-max-open-weight-flagship-first/?ref=runtimewire) that "the model is a 2.4-trillion-parameter mixture of experts activating 95 billion per token." That figure comes from a secondary account. RuntimeWire could not confirm the total or active parameter counts in official Alibaba documentation supplied for review. CNBC reported that Alibaba made the Max weights available for download and execution. The reviewed materials do not include an official Qwen3.8-Max model card or another Alibaba document that establishes its license. Developers considering commercial deployment need to verify usage rights and technical requirements independently. ### The documentation gap changes the deployment decision The August 17 release gives developers two different levels of certainty. Qwen3.8-27B comes with concrete files, Apache 2.0 terms and serving instructions. Qwen3.8-Max provides downloadable flagship parameters without equivalent license documentation in the reviewed materials. That distinction matters because "open weight" describes access to parameters, not a uniform set of rights. A model can be downloadable without granting unrestricted modification, redistribution or commercial use. Until Alibaba publishes or identifies the governing terms for Qwen3.8-Max, teams evaluating the flagship must treat licensing as an unresolved deployment dependency.