Kimi's official launch blog. Supports the 2.8-trillion-parameter architecture and benchmark disclosures plus Moonshot's promise to publish full weights and a fuller technical report by July 27; the checkpoint was still pending when this thread was posted.
Qwen's official Hugging Face organization. No Qwen3.8 model or checkpoint was visible here when checked on July 19, 2026; this supports only the time-bounded artifact check and should be revisited when weights ship.
Alibaba Cloud's official documentation. Supports the preview model's reasoning, image-understanding, and text-generation description, and states that the preview will keep changing and may later be removed or replaced.
QwenCloud's official Token Plan page. Confirms that Qwen3.8-Max-Preview is currently offered through paid hosted access rather than as a downloadable checkpoint.
Qwen's official announcement. Supports Alibaba's claims of 2.4 trillion parameters, open weights coming 'soon,' availability through Token Plan/Qoder/QoderWork, and performance 'second only to Fable 5'; the post supplies no benchmark table or weight-release date.
Sensemaker thread explaining that Qwen3.8-Max-Preview is available through Alibaba's hosted tools, while the weights, license, architecture, model card, benchmark evidence, and release date remain unpublished.
The end
