Qwen4-Max Official Release EvaluationProvider report
2026-11-16Qwen4-Max only: 16 metrics from the official 'Comprehensive Performance Comparison Across Coding, Work, and Long-Horizon Tasks' chart plus 3 items from the post body (Work360 and two Last Secure Code Bench metrics), 19 in total, all vendor-reported; the chart image SHA-256 is recorded in each score's raw. JobBench was re-run on the updated dataset and evaluation code and is stored as a separate criterion; the Chartography Qwen values were measured with min_pixels 4096x32x32, kept separate from Anthropic's tool variants; Agents' Last Exam v1.1 quotes the official leaderboard and is kept separate from Tencent's re-tested pass rate; citation sources for CodeArena/AgentArena (Arena), AA-Briefcase v1.1 and Automationbench-AA (Artificial Analysis), APEX-Agents v1.1 (Mercor) and HLAB-Vals-1.1 (Vals Index) are recorded in each score's raw; HLE keeps the default text-only metric. Comparison columns (Qwen3.8-Max-0902, Opus5.5, GPT6.1-Sol, GPT6-Astra, Kimi K3.1) are kept in raw.source_comparison and not imported as scores; the Qwen3.8-Max-0902 column matches the existing 0902 report (DeepSWE 69.3, ProgramBench 28.0). The blog page's Published metadata reads 2026-08-03 (identical to the Qwen3.8-Max release timestamp, apparently inherited from the template); the article citation states November 2026 and the official API went live on QwenCloud and OpenRouter on 2026-11-16; smaller models in the series are not yet open.