# Benchmark settings and caveats: https://huggingface.co/zai-org/GLM-5.3#footnotes
name = "GLM-5.3"
description = "Flagship GLM model for long-horizon coding, agents, and complex project delivery"
family = "glm"
release_date = "2026-08-14"
last_updated = "2026-08-14"
attachment = false
reasoning = true
temperature = true
tool_call = true
structured_output = true
open_weights = true

[limit]
context = 1_000_000
output = 131_072

[modalities]
input = ["text"]
output = ["text"]

[[benchmarks]]
name = "Terminal-Bench"
score = 88.2
metric = "pass@1"
variant = "max effort"
version = "2.1"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "Terminal-Bench"
score = 28.3
metric = "avg@3"
variant = "max effort"
version = "3.0"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "DeepSWE"
score = 66.9
metric = "resolved"
variant = "max effort; 6h timeout; 400K context"
version = "1.1"
source = "https://huggingface.co/zai-org/GLM-5.3"
harness = "mini-swe-agent"

[[benchmarks]]
name = "NL2Repo"
score = 58.0
metric = "score"
variant = "max effort"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "ProgramBench"
score = 19.0
metric = "almost solved"
variant = "max effort"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "FrontierSWE"
score = 78.1
metric = "dominance score"
variant = "max effort"
date = "2026-08-14"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "SWE-Marathon"
score = 42.5
metric = "score"
variant = "max effort; modified anti-cheat checks"
version = "1.1"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "PostTrainBench"
score = 39.8
metric = "weighted average"
variant = "max effort; 3 runs; modified external-API checks"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "CyberGym"
score = 84.5
metric = "pass@1"
variant = "max effort"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "ExploitGym"
score = 105
metric = "tasks solved"
variant = "max effort; 2h rescaled timeout"
dataset = "869 tasks"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "ExploitGym"
score = 130
metric = "tasks solved"
variant = "max effort; 6h rescaled timeout"
dataset = "869 tasks"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "ExploitBench"
score = 54.4
metric = "average coverage"
variant = "max effort; union over 3 revisions"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "Toolathlon-Verified"
score = 73
metric = "pass@1"
variant = "max effort"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "AutomationBench"
score = 48.2
metric = "pass@1"
variant = "max effort"
version = "1.0.6"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "Agents' Last Exam"
score = 28.5
metric = "score"
variant = "max effort; CLI"
harness = "Claude Code 2.1.207"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 62.5
metric = "accuracy"
variant = "max effort; with tools"
source = "https://huggingface.co/zai-org/GLM-5.3"

[[benchmarks]]
name = "GDPval-AA"
score = 1769
metric = "Elo"
variant = "max effort"
version = "2"
source = "https://huggingface.co/zai-org/GLM-5.3"