Qwen3.8-Max: 2.4 Trillion Parameter Monster

On August 3, 2026, Alibaba's Qwen team released Qwen3.8-Max, a large language model with 2.4 trillion parameters, though only about 95 billion are active per request due to its Mixture-of-Experts architecture. It has a one-million-token context window and is available via QwenCloud, Alibaba Cloud Model Studio, and the QwenWork app. The company plans to release open weights for Qwen3.8-Max and the smaller Qwen3.8-27B on August 12, marking the first time a Max-class model's weights will be open. Demonstrations show the model autonomously building a self-updating CLI tool over 16 days, reproducing and improving a scientific paper, competing in a real competition, designing a cryptographic chip, and trading in a simulated e-commerce environment. Independent testing by Artificial Analysis gave it 56 points on the Intelligence Index, one point behind Kimi K3, with a higher cost per task. The company attributed lower initial scores to API overload. Benchmarks show mixed results: it leads on PaperBench, IFBench, and OSWorld-Verified, but trails on SWE-bench Pro and FrontierSWE. Pricing is $2 per million input tokens and $6 per million output tokens, cheaper than many competitors. The community looks forward to the compact 27B model runnable on consumer hardware. The release highlights a trend where the harness (or infrastructure) around the model is increasingly important, and Alibaba claims compatibility with various harnesses like Claude Code and Codex.