Qwen3.8 Max

Shattered.io's September 7, 2026 analysis covers Alibaba's open-weight release of Qwen3.8-Max, packaged on Hugging Face and ModelScope as Qwen3.8-2.4T-A95B, after the model first became available through Alibaba's QwenCloud API on August 2, 2026. The piece notes the open-weight variant landed on August 13 and a dense s The article highlights the architectural design that makes self-hosting feasible: a 2.4 trillion-parameter sparse mixture-of-experts model with only 95 billion active parameters per forward pass. Alibaba's Qwen.ai blog positioned the release under the headline "A New Bar for Coding and Cowork," targeting software engin

DataCamp reports that Alibaba released Qwen3.8-Max-0902 on September 2, 2026, a post-training upgrade to Qwen3.8-Max focused on coding and agent performance. The architecture is unchanged from the base model: 2.4 trillion parameters and a 1M token context window. Pricing also remains unchanged at $2 per 1M input tokens All 8 coding benchmarks improved over the base Qwen3.8-Max, with the largest gains on TerminalBench 3.0 (11.3 to 29.0) and ProgramBench Almost Solved (10.5 to 28.0), both more than doubling. Qwen3.8-Max-0902 ranks first overall on Code Arena WebDev with 1,691 points, and leads Claude Opus 5 Max on MLS-Bench-Lite, SWE-A

The MindStudio article provides a third-party architecture overview of Qwen3.8-2.4T-A95B, the open-weight designation for the Qwen3.8-Max flagship, confirming the 2.4 trillion total parameter count and the 95 billion activated parameter MoE configuration announced by Alibaba. It frames the model as a major MoE release By covering the 2.4T-A95B variant explicitly rather than the 27B sibling, the article supplies architecture-level corroboration for the Qwen3.8-Max subject independent of Alibaba's own announcement. It draws on the official model card and launch artifacts to summarize design choices and reported benchmark results, givi

BenchLM's third-party dashboard for Qwen3.8-Max aggregates 60 published benchmark rows with explicit verification flags: Agentic 15/15, Coding 12/12, Reasoning 2/2, Knowledge 6/6, and Multimodal 24/24 are all verified, while Math and Multilingual are not measured. The model achieves a Capability composite of 71.8/100, The dashboard reports a 1M token context window matching DataCamp's figure, measured speed of approximately 41 tokens per second (field median 91 tok/s), and a first-token latency of 51.78 seconds. However, BenchLM explicitly notes that no first-party API token rate has been published for Qwen3.8-Max, limiting its usef

On August 3, 2026, Alibaba officially announced Qwen3.8-Max as the most powerful model in its Qwen series to date, featuring 2.4 trillion total parameters with a Sparse Mixture-of-Experts architecture that activates just 95 billion parameters, combined with a hybrid attention mechanism. The model supports a context win Qwen3.8-Max ranks fourth in the Fronted Code Arena and is described as demonstrating exceptional proficiency in autonomous coding and long-horizon execution, operating independently over extended periods without human intervention. The architecture builds on Qwen 3.5 and is designed to balance massive scale with infere

On July 19, 2026, during the World AI Conference in Shanghai, Alibaba's Qwen team previewed Qwen3.8-Max-Preview, described in the announcement as a 2.4 trillion-parameter multimodal model positioned as the next flagship in the Qwen family. The preview arrived two days after Moonshot AI released its 2.8 trillion-paramet The MarkTechPost article notes that the preview build is purchasable through Alibaba's Token Plan subscription at 10% of standard pricing, but flags that the benchmark table, model card, and license were not yet available alongside the preview. It separates what Alibaba confirmed from what was only claimed, and reports