Atlas

Models

← All models

GLM-5

Zhipu AI (Z.ai) · 2026-02-12 · 753.9B parameters

GLM-5 is Z.ai's February 12, 2026 open-weight MoE language model for complex systems engineering and long-horizon agentic tasks. Its released checkpoint uses 78 transformer layers, MLA with DeepSeek Sparse Attention, a 202,752-token context window, and turn-level thinking controls; Z.ai describes it as 744B total/40B active and reports 28.5T tokens across base-model pre-training and mid-training. The weights are released under the MIT license.

Benchmark scores

BenchmarkScore
A-Fantasia66.7
A-Fantasia Backwards Spelling78.0
A-Fantasia Chess52.0
A-Fantasia Cube Rotation70.0
AA-LCR63.3
AA-Omniscience Index2.0
AHK-Eval80.6
AHK-Eval Algorithms83.3
AHK-Eval Data Structures66.7
AHK-Eval Date and Time83.3
AHK-Eval Easy83.3
AHK-Eval Hard66.7
AHK-Eval Hidden Cases71.8
AHK-Eval Mid58.3
AHK-Eval Numbers83.3
AHK-Eval Parse Success97.2
AHK-Eval Regex50.0
AHK-Eval Strings50.0
AIME 2024 (avg@8)93.3
AIME 2024-202591.7
AIME 202596.7
AIME 2025 (avg@8)90.0
AIME 202696.7
ALE-Bench1194
Apex (MathArena)10.9
APEX (Mercor)7.3
Apex Shortlist68.6
APEX-Agents17.2
APEX-Agents-AA14.5
APEX-v1-extended49.0
Loading Atlas data…