Access and disclosure diverge in open-weight releases
Qwen3.8-Max is available through interfaces while its weights are deferred; H3's weights are in a repository without a benchmark table. The releases show why openness is not a single yes-or-no property.
Artificial Intelligence··Morning
Qwen3.8-Max access opened while the weights wait
Alibaba opened Qwen3.8-Max to global developers through Alibaba Cloud Model Studio interfaces and QwenWork, which entered public beta on the same day. The company account carried by the South China Morning Post describes a model with 2.4 trillion parameters with a one-million-token context window and says open weights will follow the next week. That timetable separates service access available today from model weights promised for later download. Alibaba says the model can build searchable knowledge bases from material ranging from long documents to television series and live streams, recreate applications from screenshots, generate interactive games and educational animations, and turn two-dimensional floor plans into three-dimensional visualisations. These are the company's product descriptions, and no independent measurement was published beside the announcement. QwenWork is also positioned against Tencent's WorkBuddy and Moonshot AI's Kimi Work in China, and against Claude Cowork and ChatGPT Work abroad. Alibaba's ownership of the South China Morning Post makes the attribution of those capability statements particularly relevant. Today's access establishes that the hosted tool can be tried. The contents of the weights and the requirements for local deployment cannot be examined at the same level until the promised release occurs.[1]
H3 supplies downloadable files but no comparative measurement
The weights for MiniMax H3 were published in a Hugging Face repository on 3 August. The repository history shows an initial release at 02.34 UTC and diffusers-format weights added at 04.50. Its model card gives the H3-Omni-Transformer component as 33 billion parameters and says roughly 13 billion sit in AdaLN-related branches that do not need to be loaded for inference. Alongside text, the model accepts at most nine images, three video clips or three audio clips, with a total file limit of 12. Output is described as four to 15 seconds long, with a default 768-pixel short side, an option to raise resolution to 2K and a frame rate of 24 per second. It generates its own 32 kHz stereo audio, and its text encoder is based on Qwen3-VL-32B weights. The usage terms of the MiniMax H3 Community License Agreement say content passes through automated moderation and that output suspected of being unlawful, pornographic or infringing third-party rights may be blocked. Although the files can be downloaded and the component breakdown inspected, the model card contains no benchmark table. This release therefore supplies a different kind of detail about access and structure from the Qwen announcement, but it does not supply a comparative performance result.[2]
Openness is measured through several separate questions
The releases widen access on the same day but do not make the same objects verifiable. Qwen3.8-Max can be used through company interfaces, while its weights are promised for the following week. H3's weight files and model card are visible now; the card lists component size, branches not loaded for inference, input limits, output format and licence conditions. Yet H3 has no benchmark table, and no independent measurement accompanies Qwen's capability list. Their parameter figures cannot support a ranking: Qwen is described by a total model count, whereas H3 distinguishes a component and a portion not required during inference. The useful comparison is therefore the disclosure stage. Hosted access lets someone try a model; downloadable files permit inspection of architecture and local requirements; performance measurements permit same-task comparison. The announcements offer different combinations of the first two and leave the third incomplete. Open weights alone consequently does not answer questions about timing, licence, loading requirements or measured quality; the material published for each question must be assessed separately.[1], [2]
Related columns
For more information on this topic, you can read the related columns.
