Qwen 3.8 Max 2.44 trillion parameters

Alibaba Cloud’s Qwen 3.8 Max is being presented in August 2026 as the company’s newest flagship model, with outside writeups describing it as a 2.4 trillion-parameter sparse mixture-of-experts system and Alibaba’s own product pages listing it in the company’s Token Plan. AiCybr Blog said the model, identified as `qwen3.8-max-preview`, was officially unveiled on August 3 after a July preview. Geeky Gadgets, in an Aug. 6 article, described the model as a 2.44 trillion-parameter open-weight multimodal release and pointed readers to benchmark tables compiled elsewhere. Alibaba Cloud’s own public pages visible on Aug. 6 confirm current access to Qwen3.8-Max through its Token Plan, but do not in the material reviewed here publish the full benchmark table carried by secondary sites. ### Where is Qwen 3.8 Max actually available right now? Alibaba Cloud’s Token Plan page lists “the newest Qwen3.8-Max” among models customers can try, and its Model Studio landing page says the plan comes with “exclusive access to Qwen3.8-Max.” Those pages position the model as available through Alibaba’s hosted platform rather than as a general public download page in the material reviewed here. AiCybr likewise described current access as Alibaba Cloud Model Studio Token Plan access for the preview endpoint. (aicybr.com) AiCybr’s Aug. 4 guide said the model supports text, images and video as input, text output, and a 1 million-token context window. The same guide described the architecture as a sparse Mixture-of-Experts model with roughly 95 billion active parameters per token. Those details were attributed by AiCybr to Alibaba launch materials and related documentation. ### Why are there two different parameter numbers? Geeky Gadgets on Aug. 6 described Qwen 3.8 Max as a 2.44 trillion-parameter model. (alibabacloud.com) AiCybr’s Aug. 4 guide described it as a 2.4 trillion-parameter model. The difference appears to be a rounding difference across secondary reports rather than a dispute over the model’s scale, based on the material reviewed. AiCybr also said Qwen 3.8 Max is a sparse model, meaning the headline parameter count is not the same thing as the number of parameters active on each token. (aicybr.com) Its guide put active parameters at about 95 billion per token. ### What benchmark numbers are being cited most often? AiCybr’s compilation, published Aug. 4 and updated with research checked Aug. 5, listed several Alibaba-reported scores for Qwen 3.8 Max. (geeky-gadgets.com) The figures included 92.6% on GPQA Diamond, 67.7% on SWE-bench Pro, 86.6% on Terminal-Bench 2.1, 86.1% on OSWorld-Verified, 93.0% on PaperBench and 82.8% on IFBench. AiCybr said vendor-reported numbers were labeled as such and distinguished from third-party scores where available. (aicybr.com) AiCybr also cautioned that independent verification was still limited because the model had been released only days earlier. The site recommended cross-checking live leaderboards and third-party evaluation platforms as more testing appears. ### Has Alibaba itself published the same benchmark table publicly? Alibaba Cloud’s public Token Plan and Model Studio landing pages reviewed on Aug. 6 show Qwen3.8-Max as a current product, but the excerpts available there do not include the detailed benchmark table cited by AiCybr. (aicybr.com) That means the headline performance claims now circulating are, in this reporting, coming through secondary compilations that say they are based on Alibaba launch materials. Geeky Gadgets’ Aug. 6 article linked the model description to benchmark discussion, helping spread the scorecard beyond Alibaba’s own ecosystem. But the benchmark figures themselves remain described by AiCybr as Alibaba-reported numbers pending broader outside testing. ### What should readers watch next? August 2026 is the next checkpoint for this story because Alibaba Cloud is already listing Qwen3.8-Max in Token Plan access, while AiCybr said independent evaluations are still emerging. (alibabacloud.com) The near-term evidence to watch is whether Alibaba publishes fuller primary materials and whether third-party benchmark platforms post repeatable Qwen 3.8 Max results under the `qwen3.8-max-preview` model identifier. (geeky-gadgets.com)