阿里Qwen3.8正式发布，编程与办公再进化，推理更快更稳定

TL;DR - Alibaba released Qwen3.8, a 2.4T-parameter (95B active) sparse-MoE flagship with 1M-token context and vision support, positioned as a top-tier model for autonomous coding and long-horizon professional "Cowork" tasks. It matters because it pairs frontier agentic benchmark claims with aggressive pricing (40%/24% of Opus 5 input/output internationally) and a promised open-source release.

• Architecture: sparse MoE plus hybrid attention scales total params to 2.4T with 95B activated, targeting faster/cheaper inference; 1M-token context and native visual understanding. Alibaba's Zhenwu M890 supernode claims up to 1.5x speedup in agentic inference.
• Claimed benchmarks: PaperBench 93.0 (+28.2 over prior gen), WideSearch 81.9, Agent's Last Exam 52.4, IFBench 82.8, GPQA Diamond 92.6, OSWorld-Verified 86.1 (reported first among mainstream models), BabyVision 82.0 tool-free; 4th on CodeArena, 2nd on Vision Arena, and second only to Claude on Arena overall.
• Agentic showcase: ran ~16 days unattended from an empty folder to produce "oh-my-cli," an open-sourced self-evolving agent framework, plus a RecreationBench app-replication task solved offline via iterative visual coding.
• Availability: API live on the Qwen AI platform (¥12/M input, ¥36/M output, ¥1.5 cached) and wired into the new "Qianwen Office" agent product; Qwen3.8-Max and Qwen3.8-27B slated to open-source next week.