GLM-4.5
Active Parameters
355B
Context Length
128K
Modality
Multimodal
Architecture
Mixture of Experts (MoE)
License
MIT License
Release Date
28 Jul 2025
Knowledge Cutoff
Jan 2025
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
41x RTX 4090
24GB VRAM
Datacenter
11x NVIDIA A100
80GB VRAM
Apple Silicon
9x Apple M3 Max
128GB VRAM
128,000 tokens
Consumer
45x RTX 4090
24GB VRAM
Datacenter
12x NVIDIA A100
80GB VRAM
Apple Silicon
9x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#91
| Benchmark | Score | Rank |
|---|---|---|
Graduate-Level QA | 0.791 | 30 |
Professional Knowledge | 0.81 | 32 |
General Text | 1411 | 84 |
Rankings
Overall Rank
#91
Coding Rank
-
About GLM-4.5
GLM-4.5 is a flagship multimodal large language model developed by Z.ai that integrates complex reasoning, software engineering, and agentic capabilities within a unified architecture. It employs a sophisticated Mixture-of-Experts (MoE) design with 355 billion total parameters, specifically engineered to optimize parameter efficiency by activating only 32 billion parameters during a forward pass. A defining feature of the model is its dual-mode execution framework, which allows it to alternate between a high-latency 'Thinking Mode' for multi-step planning and an instantaneous 'Non-Thinking Mode' for standard interactive tasks.
Technical innovations in GLM-4.5 focus on architectural depth over width to enhance logical deduction and mathematical processing. The model utilizes Grouped-Query Attention (GQA) with 96 attention heads and a hidden dimension size of 5120. Its MoE implementation features sigmoid-gated routing and QK-Norm to ensure stable expert utilization and load balancing. The training pipeline involved a massive 23-trillion-token corpus, including 7 trillion tokens dedicated to code and reasoning datasets, followed by reinforcement learning using the custom-built 'slime' infrastructure to refine autonomous decision-making.
Designed for production-grade agent applications, GLM-4.5 supports native function calling and complex web browsing with a high success rate. It features an expansive 128,000-token context window and a substantial maximum output limit of 96,000 tokens, making it suitable for long-form document analysis and full-stack software development. The model is released with open weights under the MIT License, facilitating broad adoption in both research and commercial environments.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
96
Key-Value Heads
8
Attention Head Dimension
128
Position Embedding
Absolute Position Embedding
RoPE Theta
1,000,000
Sliding Window Attention
No
Sliding Window Size
-
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
5,120
Number of Layers
96
FFN Intermediate Size (Dense)
1,536
Multi-Token Prediction Heads
1
Tokenizer
Vocabulary Size
151,552
Mixture of Experts
Total Expert Parameters
32.0B
Number of Experts
160
Active Experts
8
Shared Experts
1
FFN Intermediate Size (per Expert)
1,536
Dense Layers Before MoE
3
About GLM Family
General Language Models from Z.ai
