Skip to content
AI Atlas
Family

GLM4.5

All 4 models in the terminal →

data quality35

first seen 12 Sept 2026

family_01M2AS5GP7JWZWFJHTKCCX1BYV

Models
4
First release
20 Jul 2025
Latest release
20 Oct 2025
Parameters
81.9B – 358.3B
Artifacts
1
Providers
2

Releases

Family releases by statusActiveDeprecatedAug 25Sept 25Oct 25GLM 4.5 — 20 Jul 2025GLM 4.5 Air — 20 Jul 2025GLM 4.5V — 11 Aug 2025cerebras/GLM-4.5-Air-REAP-82B-A12B — 20 Oct 2025

4 dated releases · dot size = best benchmark rank of the member

Members 4

Family members
ModelParamsContextOpennessLicenceReleasedStatusBest rankCompare
cerebras/GLM-4.5-Air-REAP-82B-A12BCerebras Systems81.9B12B active—Open weightsMIT20 Oct 2025Active—
GLM 4.5VZ.ai (Zhipu AI)—65.5KOpen weights—11 Aug 2025Deprecated#129 MMMU-Pro
GLM 4.5Z.ai (Zhipu AI)358.3B131.1KOpen weightsMIT20 Jul 2025Deprecated#27 SWE-bench Verified
GLM 4.5 AirZ.ai (Zhipu AI)110.5B131.1KOpen weightsMIT20 Jul 2025Deprecated#121 Terminal-Bench

Lineage among members 1 relations

Open in Graph →
Lineage of the GLM4.5 familycerebras/GLM-4.5-Air-REAP-82B-A12B fine-tuned from GLM 4.5 AirGLM 4.5 Air110.5BGLM 4.5 Aircerebras/GLM-4.5-Air-REAP…81.9Bcerebras/GLM-4.5-Air-REAP-82B-A12B

Benchmark progress

Terminal-Benchbest #115

Best rank of the family on Terminal-Bench#150#100#50#0Jul 25Jul 25Aug 25Aug 25

τ²-benchbest #151

Best rank of the family on τ²-bench#200#150#100#50#0Jul 25Jul 25Aug 25Aug 25

GPQA Diamondbest #154

Best rank of the family on GPQA Diamond#200#150#100#50#0Jul 25Jul 25Aug 25Aug 25

Humanity's Last Exambest #161

Best rank of the family on Humanity's Last Exam#200#150#100#50#0Jul 25Jul 25Aug 25Aug 25

IFBenchbest #174

Best rank of the family on IFBench#250#200#150#100#50#0Jul 25Jul 25Aug 25Aug 25

Artificial Analysis Intelligence Indexbest #188

Best rank of the family on Artificial Analysis Intelligence Index#250#200#150#100#50#0Jul 25Jul 25Aug 25Aug 25

Best rank (primary comparability group, current leaderboard) reached by any member released up to each date — lower is better, axis inverted. Ranks are today's ranks, not the ranks at release time.

Aggregates are computed from the members' sourced claims; members without a release date or parameter count are left out of the range and timeline rather than guessed. Methodology →