模型能力对比(24 个模型)

测试集:根据本人实际工作使用习惯设计 · 107 题 · 8 维度 · 生成 2026-10-29 23:49 · 离线单文件

① 总榜(按综合分降序)

模型名次综合分95% CI档位一票否决剔除空答后
DeepSeek-V4.1-Flash-Fast (CommandCode)思考强度 默认10.90020.8445 – 0.9488D⛔ d7-04d7-020.9106
LongCat-2.5-Preview-Free思考强度 默认20.89790.8475 – 0.9417D⛔ d7-04d7-02d8-020.9156
Qwen3.8-Max-0902 (百炼 Token Plan)思考强度 默认30.89340.8346 – 0.9429D⛔ d7-04d7-020.8934
DeepSeek-V4.1-Flash-1017思考强度 默认40.89300.8344 – 0.9393Dd7-02d8-020.9000
DeepSeek-V4.1-Flash (CommandCode)思考强度 默认50.89220.8287 – 0.9446Dd7-020.9027
Step-5-Preview思考强度 默认60.88930.8280 – 0.9412D⛔ d7-04d7-020.9055
qwen3.8-flash思考强度 默认70.88670.8266 – 0.9385D⛔ d7-04d7-020.8867
deepseek-v4.1-flash-high思考强度 high80.88380.8229 – 0.9367D⛔ d7-04d7-020.8908
MiMo-V2.6-Pro思考强度 默认90.88290.8313 – 0.9286D⛔ d7-04d7-02d8-020.9077
LongCat-2.5-Preview-Free-high思考强度 high100.88270.8217 – 0.9344D⛔ d7-04d7-02d8-020.8941
Ling-3.1-Flash (CommandCode)思考强度 默认110.88030.8223 – 0.9287D⛔ d7-04d7-020.8896
GPT-6-Luna (CommandCode)思考强度 默认120.87380.8079 – 0.9292D⛔ d7-04d7-02d8-020.8738
MiniMax-M3.1-Flash-Free思考强度 默认130.87160.8026 – 0.9245D⛔ d1-02⛔ d7-04⛔ d7-05d7-020.9034
MiniMax-M3.1-Flash-Free-high思考强度 high140.86660.8044 – 0.9206D⛔ d7-04⛔ d7-05d7-020.8701
Kimi-K2.6 (CommandCode)思考强度 默认150.86430.8021 – 0.9184D⛔ d7-04d7-020.8903
deepseek-v4.1-flash基准思考强度 默认160.86240.7955 – 0.9209Dd7-020.8624
Claude-Haiku-5.5-medium思考强度 medium170.84920.7787 – 0.9104D⛔ d7-04d7-020.8492
Qwen3.8-27B (CommandCode)思考强度 默认180.84540.7774 – 0.9052D⛔ d7-04d7-020.8970
Qwen3.6-35B-A3B思考强度 默认190.82560.7408 – 0.9016D⛔ d1-02d7-020.8427
LongCat-2.0 (CommandCode)思考强度 默认200.82220.7584 – 0.8802D⛔ d7-04d7-020.8379
Iceman-Contributor-Free思考强度 默认210.80640.7381 – 0.8713D⛔ d7-04d7-02d8-020.8064
Ornith-1.5-35B-A3B思考强度 默认220.79410.7091 – 0.8739Dd7-020.8812
GLM-5.3-Flash思考强度 默认230.79380.7107 – 0.8674D⛔ d7-04d7-020.8178
GLM-5.3-Flash-high思考强度 high240.78330.6952 – 0.8630D⛔ d7-05d7-020.7996
名次按综合分从高到低排(综合分相同则同名次);一票否决失守会直接压档(与总分无关)。

② 八维通过率(列按权重降序)

模型D2 工具协议
20%
D1 规则遵循
18%
D3 代码
16%
D4 数据一致性
14%
D5 中文写作
12%
D6 长上下文
8%
D7 元认知
8%
D8 知识不幻觉
4%
加权总分
DeepSeek-V4.1-Flash-Fast (CommandCode)思考强度 默认79.2n=4890.9n=3397.8n=45100.0n=3693.3n=30100.0n=3062.2n=45100.0n=360.9002
LongCat-2.5-Preview-Free思考强度 默认87.5n=4893.9n=3388.9n=45100.0n=3686.7n=30100.0n=3062.2n=4594.4n=360.8979
Qwen3.8-Max-0902 (百炼 Token Plan)思考强度 默认77.1n=4890.9n=33100.0n=45100.0n=3693.3n=3096.7n=3057.8n=45100.0n=360.8934
DeepSeek-V4.1-Flash-1017思考强度 默认81.2n=4890.9n=3395.6n=45100.0n=3690.0n=30100.0n=3064.4n=4586.1n=360.8930
DeepSeek-V4.1-Flash (CommandCode)思考强度 默认79.2n=4890.9n=3397.8n=45100.0n=3686.7n=30100.0n=3062.2n=45100.0n=360.8922
Step-5-Preview思考强度 默认85.4n=48100.0n=3388.9n=45100.0n=3673.3n=30100.0n=3064.4n=4591.7n=360.8893
qwen3.8-flash思考强度 默认77.1n=4890.9n=3397.8n=45100.0n=3693.3n=3086.7n=3066.7n=4594.4n=360.8867
deepseek-v4.1-flash-high思考强度 high77.1n=4890.9n=3393.3n=45100.0n=3690.0n=30100.0n=3062.2n=4597.2n=360.8838
MiMo-V2.6-Pro思考强度 默认87.5n=4897.0n=3386.7n=4594.4n=3683.3n=3083.3n=3071.1n=4597.2n=360.8829
LongCat-2.5-Preview-Free-high思考强度 high85.4n=4890.9n=3388.9n=45100.0n=3683.3n=3096.7n=3062.2n=4597.2n=360.8827
Ling-3.1-Flash (CommandCode)思考强度 默认85.4n=4890.9n=3397.8n=45100.0n=3676.7n=3093.3n=3053.3n=45100.0n=270.8803
GPT-6-Luna (CommandCode)思考强度 默认87.5n=4890.9n=3386.7n=45100.0n=3683.3n=30100.0n=3051.1n=4588.9n=360.8738
MiniMax-M3.1-Flash-Free思考强度 默认81.2n=4887.9n=3391.1n=45100.0n=3693.3n=30100.0n=3051.1n=4580.6n=360.8716
MiniMax-M3.1-Flash-Free-high思考强度 high81.2n=4893.9n=3391.1n=4597.2n=3686.7n=3096.7n=3046.7n=4586.1n=360.8666
Kimi-K2.6 (CommandCode)思考强度 默认81.2n=4890.9n=3393.3n=4597.2n=3680.0n=3090.0n=3057.8n=4596.3n=270.8643
deepseek-v4.1-flash基准思考强度 默认87.5n=48100.0n=3391.1n=4591.7n=3663.3n=3080.0n=3066.7n=45100.0n=360.8624
Claude-Haiku-5.5-medium思考强度 medium81.2n=4878.8n=3391.1n=45100.0n=3683.3n=3093.3n=3055.6n=45100.0n=360.8492
Qwen3.8-27B (CommandCode)思考强度 默认85.4n=4890.9n=3384.4n=45100.0n=3673.3n=3083.3n=3055.6n=4591.7n=360.8454
Qwen3.6-35B-A3B思考强度 默认87.5n=4863.6n=3380.0n=45100.0n=3690.0n=3090.0n=3060.0n=45100.0n=270.8256
LongCat-2.0 (CommandCode)思考强度 默认83.3n=4884.8n=3384.4n=4591.7n=3673.3n=3090.0n=3051.1n=4596.3n=270.8222
Iceman-Contributor-Free思考强度 默认85.4n=48100.0n=3368.9n=4591.7n=3650.0n=30100.0n=3055.6n=4581.5n=270.8064
Ornith-1.5-35B-A3B思考强度 默认68.8n=4881.8n=3386.7n=45100.0n=3650.0n=3090.0n=3073.3n=45100.0n=360.7941
GLM-5.3-Flash思考强度 默认83.3n=4866.7n=3377.8n=4588.9n=3676.7n=30100.0n=3057.8n=45100.0n=360.7938
GLM-5.3-Flash-high思考强度 high72.9n=4872.7n=3388.9n=4580.6n=3666.7n=30100.0n=3064.4n=45100.0n=360.7833
底色深浅 = 通过率:0%25%50%75%100%
⚠ 分母不同,这些维度不要横向比:D8(DeepSeek-V4.1-Flash-Fast (CommandCode) n=36 / LongCat-2.5-Preview-Free n=36 / Qwen3.8-Max-0902 (百炼 Token Plan) n=36 / DeepSeek-V4.1-Flash-1017 n=36 / DeepSeek-V4.1-Flash (CommandCode) n=36 / Step-5-Preview n=36 / qwen3.8-flash n=36 / deepseek-v4.1-flash-high n=36 / MiMo-V2.6-Pro n=36 / LongCat-2.5-Preview-Free-high n=36 / Ling-3.1-Flash (CommandCode) n=27 / GPT-6-Luna (CommandCode) n=36 / MiniMax-M3.1-Flash-Free n=36 / MiniMax-M3.1-Flash-Free-high n=36 / Kimi-K2.6 (CommandCode) n=27 / deepseek-v4.1-flash n=36 / Claude-Haiku-5.5-medium n=36 / Qwen3.8-27B (CommandCode) n=36 / Qwen3.6-35B-A3B n=27 / LongCat-2.0 (CommandCode) n=27 / Iceman-Contributor-Free n=27 / Ornith-1.5-35B-A3B n=36 / GLM-5.3-Flash n=36 / GLM-5.3-Flash-high n=36) —— 「未参评」= 模型不具备该能力(如纯文本模型没有视觉题),不计分母不扣分。

③ 八维点图(点=通过率,横线=95% CI)

D2 工具协议 ·20% ·48 题0255075100DeepSeek-V4.1-Fl…79.2LongCat-2.5-Prev…87.5Qwen3.8-Max-0902…77.1DeepSeek-V4.1-Fl…81.2DeepSeek-V4.1-Fl…79.2Step-5-Preview85.4qwen3.8-flash77.1deepseek-v4.1-fl…77.1MiMo-V2.6-Pro87.5LongCat-2.5-Prev…85.4Ling-3.1-Flash (…85.4GPT-6-Luna (Comm…87.5MiniMax-M3.1-Fla…81.2MiniMax-M3.1-Fla…81.2Kimi-K2.6 (Comma…81.2deepseek-v4.1-fl…87.5Claude-Haiku-5.5…81.2Qwen3.8-27B (Com…85.4Qwen3.6-35B-A3B87.5LongCat-2.0 (Com…83.3Iceman-Contribut…85.4Ornith-1.5-35B-A…68.8GLM-5.3-Flash83.3GLM-5.3-Flash-hi…72.9
D1 规则遵循 ·18% ·33 题0255075100DeepSeek-V4.1-Fl…90.9LongCat-2.5-Prev…93.9Qwen3.8-Max-0902…90.9DeepSeek-V4.1-Fl…90.9DeepSeek-V4.1-Fl…90.9Step-5-Preview100.0qwen3.8-flash90.9deepseek-v4.1-fl…90.9MiMo-V2.6-Pro97.0LongCat-2.5-Prev…90.9Ling-3.1-Flash (…90.9GPT-6-Luna (Comm…90.9MiniMax-M3.1-Fla…87.9MiniMax-M3.1-Fla…93.9Kimi-K2.6 (Comma…90.9deepseek-v4.1-fl…100.0Claude-Haiku-5.5…78.8Qwen3.8-27B (Com…90.9Qwen3.6-35B-A3B63.6LongCat-2.0 (Com…84.8Iceman-Contribut…100.0Ornith-1.5-35B-A…81.8GLM-5.3-Flash66.7GLM-5.3-Flash-hi…72.7
D3 代码 ·16% ·45 题0255075100DeepSeek-V4.1-Fl…97.8LongCat-2.5-Prev…88.9Qwen3.8-Max-0902…100.0DeepSeek-V4.1-Fl…95.6DeepSeek-V4.1-Fl…97.8Step-5-Preview88.9qwen3.8-flash97.8deepseek-v4.1-fl…93.3MiMo-V2.6-Pro86.7LongCat-2.5-Prev…88.9Ling-3.1-Flash (…97.8GPT-6-Luna (Comm…86.7MiniMax-M3.1-Fla…91.1MiniMax-M3.1-Fla…91.1Kimi-K2.6 (Comma…93.3deepseek-v4.1-fl…91.1Claude-Haiku-5.5…91.1Qwen3.8-27B (Com…84.4Qwen3.6-35B-A3B80.0LongCat-2.0 (Com…84.4Iceman-Contribut…68.9Ornith-1.5-35B-A…86.7GLM-5.3-Flash77.8GLM-5.3-Flash-hi…88.9
D4 数据一致性 ·14% ·36 题0255075100DeepSeek-V4.1-Fl…100.0LongCat-2.5-Prev…100.0Qwen3.8-Max-0902…100.0DeepSeek-V4.1-Fl…100.0DeepSeek-V4.1-Fl…100.0Step-5-Preview100.0qwen3.8-flash100.0deepseek-v4.1-fl…100.0MiMo-V2.6-Pro94.4LongCat-2.5-Prev…100.0Ling-3.1-Flash (…100.0GPT-6-Luna (Comm…100.0MiniMax-M3.1-Fla…100.0MiniMax-M3.1-Fla…97.2Kimi-K2.6 (Comma…97.2deepseek-v4.1-fl…91.7Claude-Haiku-5.5…100.0Qwen3.8-27B (Com…100.0Qwen3.6-35B-A3B100.0LongCat-2.0 (Com…91.7Iceman-Contribut…91.7Ornith-1.5-35B-A…100.0GLM-5.3-Flash88.9GLM-5.3-Flash-hi…80.6
D5 中文写作 ·12% ·30 题0255075100DeepSeek-V4.1-Fl…93.3LongCat-2.5-Prev…86.7Qwen3.8-Max-0902…93.3DeepSeek-V4.1-Fl…90.0DeepSeek-V4.1-Fl…86.7Step-5-Preview73.3qwen3.8-flash93.3deepseek-v4.1-fl…90.0MiMo-V2.6-Pro83.3LongCat-2.5-Prev…83.3Ling-3.1-Flash (…76.7GPT-6-Luna (Comm…83.3MiniMax-M3.1-Fla…93.3MiniMax-M3.1-Fla…86.7Kimi-K2.6 (Comma…80.0deepseek-v4.1-fl…63.3Claude-Haiku-5.5…83.3Qwen3.8-27B (Com…73.3Qwen3.6-35B-A3B90.0LongCat-2.0 (Com…73.3Iceman-Contribut…50.0Ornith-1.5-35B-A…50.0GLM-5.3-Flash76.7GLM-5.3-Flash-hi…66.7
D6 长上下文 ·8% ·30 题0255075100DeepSeek-V4.1-Fl…100.0LongCat-2.5-Prev…100.0Qwen3.8-Max-0902…96.7DeepSeek-V4.1-Fl…100.0DeepSeek-V4.1-Fl…100.0Step-5-Preview100.0qwen3.8-flash86.7deepseek-v4.1-fl…100.0MiMo-V2.6-Pro83.3LongCat-2.5-Prev…96.7Ling-3.1-Flash (…93.3GPT-6-Luna (Comm…100.0MiniMax-M3.1-Fla…100.0MiniMax-M3.1-Fla…96.7Kimi-K2.6 (Comma…90.0deepseek-v4.1-fl…80.0Claude-Haiku-5.5…93.3Qwen3.8-27B (Com…83.3Qwen3.6-35B-A3B90.0LongCat-2.0 (Com…90.0Iceman-Contribut…100.0Ornith-1.5-35B-A…90.0GLM-5.3-Flash100.0GLM-5.3-Flash-hi…100.0
D7 元认知 ·8% ·45 题0255075100DeepSeek-V4.1-Fl…62.2LongCat-2.5-Prev…62.2Qwen3.8-Max-0902…57.8DeepSeek-V4.1-Fl…64.4DeepSeek-V4.1-Fl…62.2Step-5-Preview64.4qwen3.8-flash66.7deepseek-v4.1-fl…62.2MiMo-V2.6-Pro71.1LongCat-2.5-Prev…62.2Ling-3.1-Flash (…53.3GPT-6-Luna (Comm…51.1MiniMax-M3.1-Fla…51.1MiniMax-M3.1-Fla…46.7Kimi-K2.6 (Comma…57.8deepseek-v4.1-fl…66.7Claude-Haiku-5.5…55.6Qwen3.8-27B (Com…55.6Qwen3.6-35B-A3B60.0LongCat-2.0 (Com…51.1Iceman-Contribut…55.6Ornith-1.5-35B-A…73.3GLM-5.3-Flash57.8GLM-5.3-Flash-hi…64.4
D8 知识不幻觉 ·4% ·样本数不同 ⚠0255075100DeepSeek-V4.1-Fl…100.0LongCat-2.5-Prev…94.4Qwen3.8-Max-0902…100.0DeepSeek-V4.1-Fl…86.1DeepSeek-V4.1-Fl…100.0Step-5-Preview91.7qwen3.8-flash94.4deepseek-v4.1-fl…97.2MiMo-V2.6-Pro97.2LongCat-2.5-Prev…97.2Ling-3.1-Flash (…100.0GPT-6-Luna (Comm…88.9MiniMax-M3.1-Fla…80.6MiniMax-M3.1-Fla…86.1Kimi-K2.6 (Comma…96.3deepseek-v4.1-fl…100.0Claude-Haiku-5.5…100.0Qwen3.8-27B (Com…91.7Qwen3.6-35B-A3B100.0LongCat-2.0 (Com…96.3Iceman-Contribut…81.5Ornith-1.5-35B-A…100.0GLM-5.3-Flash100.0GLM-5.3-Flash-hi…100.0

④ 失败构成 · 难度分层 · 速度

模型真错答空答未参评L2 通过率L3 通过率平均耗时中位P95输出 tok/s缓存命中
DeepSeek-V4.1-Flash-Fast (CommandCode)思考强度 默认303093.3%(255)75.8%(66)8 s4 s21 s133.884.2%
LongCat-2.5-Preview-Free思考强度 默认315092.2%(255)75.8%(66)32 s14 s106 s35.185.3%
Qwen3.8-Max-0902 (百炼 Token Plan)思考强度 默认350091.8%(255)77.3%(66)60 s11 s290 s41.182.8%
DeepSeek-V4.1-Flash-1017思考强度 默认362091.0%(255)77.3%(66)8 s4 s20 s153.585.6%
DeepSeek-V4.1-Flash (CommandCode)思考强度 默认323092.5%(255)75.8%(66)12 s5 s23 s101.985.2%
Step-5-Preview思考强度 默认357089.8%(255)75.8%(66)7 s3 s25 s194.985.4%
qwen3.8-flash思考强度 默认380092.9%(255)69.7%(66)44 s10 s195 s61.183.4%
deepseek-v4.1-flash-high思考强度 high362091.8%(255)74.2%(66)15 s5 s30 s79.585.2%
MiMo-V2.6-Pro思考强度 默认327089.4%(255)81.8%(66)29 s8 s106 s40.984.6%
LongCat-2.5-Preview-Free-high思考强度 high364091.0%(255)74.2%(66)31 s13 s122 s35.592.6%
Ling-3.1-Flash (CommandCode)思考强度 默认404988.2%(246)77.3%(66)20 s7 s77 s88.899.7%
GPT-6-Luna (CommandCode)思考强度 默认470089.4%(255)69.7%(66)6 s4 s12 s48.885.0%
MiniMax-M3.1-Flash-Free思考强度 默认4011085.9%(255)77.3%(66)16 s5 s83 s93.285.7%
MiniMax-M3.1-Flash-Free-high思考强度 high521085.9%(255)74.2%(66)15 s5 s47 s58.381.3%
Kimi-K2.6 (CommandCode)思考强度 默认378989.0%(246)72.7%(66)22 s10 s85 s77.783.0%
deepseek-v4.1-flash基准思考强度 默认450089.4%(255)72.7%(66)5 s3 s14 s86.484.8%
Claude-Haiku-5.5-medium思考强度 medium490086.7%(255)77.3%(66)5 s4 s10 s124.50.0%
Qwen3.8-27B (CommandCode)思考强度 默认3716086.3%(255)72.7%(66)18 s6 s96 s120.285.3%
Qwen3.6-35B-A3B思考强度 默认456987.8%(246)68.2%(66)78 s51 s184 s30.287.3%
LongCat-2.0 (CommandCode)思考强度 默认555983.7%(246)69.7%(66)26 s9 s62 s64.485.4%
Iceman-Contributor-Free思考强度 默认670980.5%(246)71.2%(66)20 s15 s32 s34.434.4%
Ornith-1.5-35B-A3B思考强度 默认2829087.1%(255)63.6%(66)84 s19 s500 s30.085.9%
GLM-5.3-Flash思考强度 默认537084.3%(255)69.7%(66)12 s7 s29 s61.636.7%
GLM-5.3-Flash-high思考强度 high555084.7%(255)68.2%(66)14 s6 s32 s42.543.5%
真错答 = 能力边界;空答 = 思考链吃光预算(工程问题,剔除后的分数见总榜「剔除空答后」列);未参评 = 不具备该能力(不计分母不扣分)。