Qwen2.5-72B
Parameters
72B
Context Length
131K
Modality
Text
Architecture
Dense
KV Cache Rate
240 KB/tok
License
Qwen License
Release Date
19 Sept 2024
Knowledge Cutoff
Jun 2024
API Pricing (per 1M)
Input: $1.20 · Output: $1.20
Hardware & System Requirements
Precision
Context
Workload Memory: 37.90 GB
INT4 · 8K context
Base Weights: 33.53 GB
KV Cache: 1.88 GB
Runtime & Graphs: 2.50 GB
Minimum Required Hardware
1x NVIDIA A40 48GB
37.9 / 43.2 GB usable
2x RTX 4090 24GB
37.9 / 43.2 GB usable · TP 2
Apple M4 Max 128GB
37.9 / 96.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Runtime requirements
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 4.43.1
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
Architecture Diagram
Evaluation Benchmarks
Rank
#165
| Benchmark | Score | Rank |
|---|---|---|
Professional Knowledge | 0.711 | 41 |
Graduate-Level QA | 0.49 | 106 |
General Text | 1303 | 168 |
Intelligence Index | 0.08 | 263 |
StackEval Archived | 0.891 | 9 |
QA Assistant Archived | 0.935 | 12 |
Summarization Archived | 0.742 | 18 |
Rankings
Overall Rank
#165
Coding Rank
-
About Qwen2.5-72B
Qwen2.5-72B is Alibaba Cloud's flagship open-weights dense model engineered for advanced coding, mathematical reasoning, and multi-turn enterprise agents. It excels at long-form generation and structured data comprehension over a 131K context window.
Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
128
Key-Value Heads
8
Position Embedding
ROPE
RoPE Theta
1,000,000
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
FP16
Hidden Dimension Size
12,288
Number of Layers
80
FFN Intermediate Size (Dense)
29,568
Tokenizer
Vocabulary Size
152,064
About Qwen2.5
Qwen2.5 by Alibaba is a family of dense, decoder-only language models available in various sizes, with some variants utilizing Mixture-of-Experts. These models are pretrained on large-scale datasets, supporting extended context lengths and multilingual communication. The family includes specialized models for coding, mathematics, and multimodal tasks, such as vision and audio processing.