Alibaba Announces Open Model Release of Qwen3.8 With 2.4 Trillion Parameters
Alibaba's return to an open model strategy with Qwen3.8, alongside Moonshot AI's Kimi K3, signals that Chinese AI companies are competing directly with US frontier models while releasing their technology openly, creating a dynamic where open models from China may match or exceed the performance of closed US models.
Key Facts
- Alibaba's Qwen3.8 has 2.4 trillion parameters and will be released as an open model.
- A preview version Qwen3.8-Max-Preview is available on Qwen Studio and Qwen Cloud's token plan.
- Moonshot AI's Kimi K3, released July 17, 2026, has 2.8 trillion parameters and a 1 million token context window.
- Kimi K3 demand caused Moonshot AI to temporarily suspend new subscriptions after GPU capacity was exhausted within 48 hours.
- Alibaba had previously released Qwen3.5 and Qwen3.6 as open models, but Qwen3.7 was closed; Qwen3.8 returns to an open model strategy.
Reporting from 1 source: GIGAZINE.
Alibaba announced it will release its proprietary AI model Qwen3.8 as an open model. The Qwen team stated the model has 2.4 trillion parameters and claimed its performance is "second only to Claude Fable 5" among current frontier AI models. A preview version, Qwen3.8-Max-Preview, is now available on the chat AI service Qwen Studio and through Qwen Cloud's token plan. Alibaba had previously released Qwen3.5 and Qwen3.6 as open models, but Qwen3.7 was only available as a closed model. With Qwen3.8, the company returns to an open model strategy. The company did not release specific benchmark results alongside the announcement. The preview version is expected to continue improving in performance after release, and the official version is anticipated to be more powerful. Separately, Chinese AI company Moonshot AI released its open model Kimi K3 on July 17, 2026, which has 2.8 trillion parameters and a 1 million token context window. Kimi K3 reportedly surpassed Claude Fable 5 and GPT-5.6 Sol in some benchmark tests. Demand for Kimi K3 was so high that Moonshot AI temporarily suspended new subscriptions after GPU capacity became insufficient within 48 hours of the announcement.
The Qwen team released a preview version of the model, Qwen3.8-Max-Preview, on the chat AI service Qwen Studio and through Qwen Cloud's token plan. Examples of SVG animations created with the preview model have been posted online. Cypress Frankenfeld, who created one of the images, said that entering "create an SVG of a pelican riding a bicycle" in Qwen Studio generated an SVG image animation. The preview version is expected to continue improving after release, and the official version is anticipated to be more powerful.
Alibaba had previously released Qwen3.5 and Qwen3.6 as open models, but Qwen3.7 was only available as a closed model. The earlier open models led to many quantized and fine-tuned versions by volunteers. On July 14, 2026, Bonsai 27B appeared, which reduced Qwen3.6-27B to a memory usage of 3.9GB, making it operable on an iPhone alone. Similar developments are expected for Qwen3.8.
Separately, Chinese AI company Moonshot AI released Kimi K3 as an open model on July 17, 2026. It has 2.8 trillion parameters and a 1 million token context window. Demand was so high that within 48 hours, Moonshot AI's GPU capacity became insufficient and the company temporarily suspended new subscription contracts to prioritize existing users. All plans on Moonshot AI's pricing page show "Sold out." Online, some posts suggest that "Alibaba is a major investor in Moonshot AI, and the Kimi series competes with the Qwen series for Alibaba Cloud's GPU capacity." AI researcher Nicolas Bustamante, who works at Microsoft, noted that the GPU shortage is a global problem, with shortages occurring in "land," "permits," "power," "chips," "construction," and "cooling." Foreign media outlet The Information reported, citing sources, that Moonshot AI has begun procedures to seek approval from investors for an IPO on the Hong Kong market.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.