Qwen3.8 Max

Shattered.io covers the original Qwen3.8-Max launch — Alibaba's first shipped the model via its QwenCloud API on August 2, 2026, then released the open-weight variant Qwen3.8-2.4T-A95B on Hugging Face and ModelScope on August 13, with a dense sibling Qwen3.8-27B following on August 14. The architecture is described as The article positions Qwen3.8-Max under the heading "A New Bar for Coding and Cowork," targeting software engineering and multi-step agent tasks rather than general chat. It notes the headline score of 86.6 referenced in the title and frames Alibaba's decision to ship open weights as a notable pushback against API-gate

DataCamp adds a dedicated update section for Qwen3.8-Max-0902, released September 2, 2026 as a post-training upgrade to Qwen3.8-Max with unchanged architecture (2.4T parameters, 1M context) and unchanged pricing ($2/1M input, $6/1M output). The article highlights the 0902 snapshot's first-place ranking on Code Arena We The piece tabulates eight coding benchmarks where 0902 improved over the prior Qwen3.8-Max, with the largest gains on TerminalBench 3.0 (11.3 → 29.0) and ProgramBench Almost Solved (10.5 → 28.0). It also provides a full coding/agent/multimodal comparison versus Claude Opus 5, showing 0902 leading on MLS-Bench-Lite, SWE

CellCog reports that Alibaba's Qwen team pushed an in-place upgrade to Qwen3.8-Max on September 1, 2026, named Qwen3.8-Max-0902 (API id qwen3.8-max-0902, alias qwen3.8-max-2026-09-02). It retains the same 2.4-trillion-parameter base and 1M-token context window but is further post-trained on what Alibaba calls "Coding & The article frames Qwen3.8-Max-0902 as a meaningful but bounded upgrade: a real sharpening on coding and agent-style office work without a price change, yet still trailing Opus 5 on most of the coding and office-work rows Alibaba chose to display. It explicitly identifies the snapshot as the variant now live on QwenClo

Qwen3.8-Max, Alibaba's 2.4-trillion-parameter mixture-of-experts multimodal model with 95 billion active parameters and a 1-million-token context window, was released on August 3, 2026 alongside a full vendor-published benchmark table. Reported published wins include PaperBench at 93.0, IFBench at 82.8, Terminal Bench The same benchmark table also shows losses against competitors, including HLE at 43.6 (last among four flagships) and SWE-bench Pro at 67.7 (12 points behind Fable 5). The article notes that all numbers are vendor-run and most coding rows used Anthropic's Claude Code harness, and that no independent evaluator had score

Alibaba officially announced Qwen3.8-Max on August 3, 2026 as its most capable flagship in the Qwen series. The model is built on the Qwen 3.5 foundation using a sparse Mixture-of-Experts architecture with a hybrid attention mechanism, totaling 2.4 trillion parameters while activating only 95 billion per forward pass. Qwen3.8-Max is available through APIs on Alibaba Cloud Model Studio for global developers, is accessible via Alibaba's QwenWork AI agent platform, and has open model weights scheduled for release the week following the announcement. The model targets complex real-world workflows including coding, research, and extended

Qwen3.8-Max was released on August 3, 2026 as QwenCloud's new flagship, listed in the official model catalog with a 1-million-token context window, Thinking, Function Calling, built-in tools, and Structured Output support. Qwen's release log describes it as a native vision-language, 2.4-trillion-parameter Mixture-of-Ex The article also clarifies an important naming distinction: Qwen3.8 Max is a 2026 model generation and is not the older Qwen3-8B eight-billion-parameter checkpoint, so any Qwen3-8B downloads, deployment guides, or pricing should not be confused with Qwen3.8-Max. Open weights were announced for future release but the Ev