In brief: Alibaba has unveiled Qwen3.8-Max with 2.4 trillion parameters and announced the release of open weights for next week, alongside the smaller Qwen3.8-27B.
Alibaba has introduced Qwen3.8-Max, its most powerful model to date, a model with 2.4 trillion parameters designed for coding, long-horizon agentic tasks, and multimodal reasoning. According to the announcement, the weights are set to be released openly next week, along with the smaller Qwen3.8-27B.
Following the so-called “Qwen exodus” last year, when a leadership change took place and the company shifted its focus toward closed model APIs, there were doubts as to whether the lab would remain relevant as a leading provider of open models. These doubts have been dispelled with the announcement of Qwen3.8-Max. The model has 2.4 trillion parameters and, according to observers, would currently be the strongest open model in the world — had Kimi K3 not already claimed that position. Via the API, Alibaba offers the model at 2 US dollars per million input tokens and 6 US dollars per million output tokens, with cached tokens priced at 0.25 US dollars per million. The company has committed to releasing both Qwen3.8-Max and the smaller Qwen3.8-27B as open weights.
For coding and agentic workflows, Alibaba cites several concrete benchmarks: the model reportedly built a self-developed coding harness unsupervised over the course of ten days. In a separate test, it independently reconstructed the pipeline of a research paper on data selection for LLM reasoning and, over a 125-hour iterative research loop, improved the original benchmark result by 2.71 points. At the WWW2025 Multimodal Dialogue Intent Recognition Challenge, the model competed against 526 human teams and, within 24 hours, reached the top 13 percent, outperforming 87 percent of the human teams.
Results are also cited in the area of hardware design: in a complete chip design process (GCD/RSA cryptographic accelerator) spanning RTL editing, simulation, synthesis, and physical layout, the model reduced the gate count from 8,298 to 678, achieving an 81 percent reduction in chip area while meeting timing closure at 500 MHz. In the so-called E-Commerce Bench, a simulation of 365 days of store operations, the model achieved a 4.16x return through continuous game-theoretic negotiations and inventory planning, corresponding to an account balance of 416,252 yen. In addition, the model integrates native visual feedback into planning, coding, and GUI interaction, which, according to Alibaba, enables the direct replication of applications across desktop, mobile, and web. Qwen-MM plugins were also released to extend existing agentic frameworks with multimodal capabilities.
For engineering teams that self-host models or integrate them into their own infrastructure, the announcement is relevant for two reasons: first, it shows that Chinese open-weight labs are now competing on equal footing with Western closed-source providers. Second, the announced release of the weights means that infrastructure providers such as Baseten are already preparing to support the model, as are tools like Hermes Agent and Command Code. Anyone planning production-grade coding agents or long-horizon autonomous workflows should carefully review the release of the weights and the associated benchmarks once they become available.
Source: www.latent.space · Published August 4, 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.