Home Model series Qwen VL Qwen3-VL-235B-A22B-Thinking Qwen3-VL-235B-A22B-Thinking · Arena-Hard V2 · Qwen3-VL技术报告v2;Table 6; page 20 footnote says win rates evaluated by GPT-4.1; Thinking MoE sampling settings · Scores and configurations Qwen3-VL-235B-A22B-Thinking · Arena-Hard V2 · Qwen3-VL技术报告v2;Table 6; page 20 footnote says win rates evaluated by GPT-4.1; Thinking MoE sampling settings · Scores and configurations Qwen3-VL-235B-A22B-Thinking record for Arena-Hard V2 · Qwen3-VL技术报告v2;Table 6; page 20 footnote says win rates evaluated by GPT-4.1; Thinking MoE sampling settings. 1 score record for this release, including 1 exact value (may include historical records). Includes 1 source location.
Scores and evaluation conditions Qwen3-VL-235B-A22B-Thinking · Arena-Hard V2 scores and configurations Metric and score Configuration and evaluation conditions Sources win rate 74.8 Reported 74.8 · Unit not specified Published Unconfirmed
Thinking · 报告v2文本评测 Settings Table6 Thinking MoE:temperature=0.6,top_p=0.95,top_k=20;普通最大输出32768,AIME-25/HMMT-25/LiveCodeBench v6为81920。报告第20–21页的presence penalty=1.5属于紧邻的Dense Thinking句,不能直接赋给235B MoE。 Unknowns 报告未明确235B MoE的presence penalty;不从Dense Thinking句或当前README复现设置推定。、与当前README复现out_seq_length=40960的最大输出说明不同,分别保留,不宣称同一次运行。、Table6多数行未明确计分单位,CFEval列未定义rating指标。 Tools Tool access not recorded Reasoning budget Output limit 32768 tokens Framework not recorded Protocol Thinking MoE · max_output 32768 Conditions Table 6; page 20 footnote says win rates evaluated by GPT-4.1; Thinking MoE sampling settings Thinking MoE sampling: temperature 0.6, top-p 0.95, top-k 20; presence penalty for MoE is not explicitly stated (the 1.5 clause follows the Dense Thinking settings); max output 32,768 tokens, except AIME-25, HMMT-25 and LiveCodeBench v6 use 81,920. 仅取235B-A22B-Thinking目标列,其他变体不是本结果的替代。 View this score in rankings
Benchmark versions and metrics Aliases Arena-Hard V2
win rate · Unit not specified · Ranking direction unconfirmed