ChatGLM2-6B
Parameters
6B
Context Length
33K
Modality
Text
Architecture
Dense
License
Custom License (ChatGLM2-6B License)
Release Date
25 Jun 2023
Knowledge Cutoff
-
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
1x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
32,768 tokens
Consumer
1x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#170
| Benchmark | Score | Rank |
|---|---|---|
General Text | 1024 | 165 |
Rankings
Overall Rank
#170
Coding Rank
-
About ChatGLM2-6B
ChatGLM2-6B is a bilingual large language model designed to facilitate conversational interactions in both Chinese and English. As the second iteration in the ChatGLM series developed by THUDM, it is built upon the General Language Model (GLM) framework and serves as a versatile tool for dialogue generation and cross-lingual text processing. The model is optimized for execution on consumer-grade hardware through efficient architectural choices, enabling a high degree of accessibility for developers and researchers working within hardware-constrained environments.
The architecture utilizes a dense transformer structure that incorporates several technical advancements over its predecessor. A key innovation is the adoption of Multi-Query Attention (MQA), which streamlines inference by sharing key and value heads across multiple query heads, significantly reducing the memory footprint of the KV cache. Furthermore, the model integrates Rotary Position Embeddings (RoPE) to capture token relationships and utilizes RMSNorm for improved training stability. The inclusion of FlashAttention during the pre-training phase allows the architecture to support a substantial context window, facilitating the processing of extended dialogue histories.
Operating with 6 billion parameters, ChatGLM2-6B provides a balanced profile of performance and efficiency. It was pre-trained on a diverse dataset comprising 1.4 trillion tokens and refined through human preference alignment to enhance its conversational quality. The model is particularly suited for applications such as intelligent virtual assistants and localized chatbots, where low-latency inference and bilingual proficiency are primary requirements. Its open-weights nature and support for INT4 quantization further expand its utility for local deployment and integration into specialized NLP pipelines.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
32
Key-Value Heads
2
Attention Head Dimension
128
Position Embedding
Absolute Position Embedding
RoPE Theta
-
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
4,096
Number of Layers
28
FFN Intermediate Size (Dense)
13,696
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
65,024
About ChatGLM
ChatGLM series models from Z.ai, based on GLM architecture.