Model

ChatGLM2-6B

OPEN

Zhipu AI / THUDM (Tsinghua) · ChatGLM family · released Jun 25, 2023

Second-gen 6B chat model; FlashAttention-based context extended to 32K, +23% MMLU / +571% GSM8K over gen 1, faster INT4 inference.

ReasoningCodingVisionFunction callingTool useAgentic

-219.2

Elo · rank #426

Parameters

6.24B

Active params

6.24B (dense)

Context

32K tokens

Architecture

Dense Transformer (GLM, Multi-Query Attention, FlashAttention)

License

Apache-2.0 (code); ChatGLM2-6B Model License for weights (free for research; free commercial use after registration)

Languages

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

BIG-Bench HardReasoning30.0%#134

best: ERNIE 4.5 300B-A47B · 94.3%

C-EvalKnowledge50.1%#51

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge50.0%#29

best: Doubao-1.5-Pro · 90.9%

DROPReasoning25.6#62

best: Hunyuan-T1 · 93.1

GSM8KMath28.1%#155

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning57.0%#145

best: Claude 3 Opus · 95.4%

HumanEvalCoding9.8%#185

best: Claude Opus 4.5 · 99.4%

MATH-500Math6.9%#195

best: GPT-5 · 99.4%

MMLUKnowledge45.5%#244

best: OpenAI o3 · 92.9%

PIQAReasoning69.6%#76

best: GPT-4o mini · 93.1%

Run it locally

VRAM @ Q4

5.5 GB

VRAM @ FP16

13.1 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

44.62 tok/s generating 2K tokens, measured on A100-SXM-80G (official repo); no RTX 4090 llama.cpp figure found

Quantizations

INT4 · INT8

ChatGLM family

Elo progression across releases

ChatGLM-6BMar 2023-456.9 ChatGLM2-6BJun 2023-219.2 ChatGLM3-6BOct 2023367.9

Resources

HFzai-org/chatglm2-6b GHzai-org/ChatGLM2-6B

API price weights · each benchmark row carries its own source badge (see methodology)