Win rate 40.4% Reported 40.4% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Claude Sonnet 4. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Tie rate 9.6% Reported 9.6% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Claude Sonnet 4. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Loss rate 50% Reported 50% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Claude Sonnet 4. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Win rate 53.9% Reported 53.9% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Kimi K2. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Tie rate 17.3% Reported 17.3% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Kimi K2. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Loss rate 28.8% Reported 28.8% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Kimi K2. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Win rate 80.8% Reported 80.8% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Qwen3-Coder. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Tie rate 7.7% Reported 7.7% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Qwen3-Coder. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Loss rate 11.5% Reported 11.5% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol CC-Bench: 52 programming tasks across six domains, isolated containers, multi-round human-expert evaluation; head-to-head against Qwen3-Coder. Conditions GLM-4.5; serving endpoint and precision are not specified. View this score in rankings Tool-calling success rate 90.6% Reported 90.6% · Unit % Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol Average tool-calling success rate on CC-Bench; 52 programming tasks across six domains. Conditions GLM-4.5; figure compares with Claude Sonnet 4, Kimi K2 and Qwen3-Coder. View this score in rankings Tokens per round 1388259 tokens Reported 1,388,259 · Unit tokens Published 2025-08-11
GLM-4.5 reported Settings GLM-4.5; 355B total, 32B active parameters. Unknowns API/hosted endpoint versus local inference and exact weight precision are not specified in the report. Tools Tool access not recorded Reasoning budget not recorded Framework not recorded Protocol Average input plus output tokens across multiple tool calls, without cache; CC-Bench. Conditions GLM-4.5; figure compares with Claude Sonnet 4, Kimi K2 and Qwen3-Coder. View this score in rankings