Phi-3-small
Parameters
7B
Context Length
8K
Modality
Text
Architecture
Dense
KV Cache Rate
128 KB/tok
License
MIT License
Release Date
22 Apr 2024
Knowledge Cutoff
Oct 2023
API Pricing (per 1M)
Input: $0.15 · Output: $0.60
Hardware & System Requirements
Precision
Context
Workload Memory: 6.76 GB
INT4 · 8K context
Base Weights: 3.26 GB
KV Cache: 1.00 GB
Runtime & Graphs: 2.50 GB
Minimum Required Hardware
1x NVIDIA L4 24GB
6.8 / 21.6 GB usable
1x RTX 3060 12GB
6.8 / 10.8 GB usable
Apple M4 24GB
6.8 / 18.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Runtime requirements
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 4.38.1
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
Additional required packages
Install alongside standard runtime libraries:
pip install einops flash_attn
Architecture Diagram
Evaluation Benchmarks
Rank
#241
| Benchmark | Score | Rank |
|---|---|---|
General Text | 1171 | 189 |
Rankings
Overall Rank
#241
Coding Rank
-
About Phi-3-small
Phi-3-small is a 7B parameter small language model from Microsoft engineered for commercial reasoning, mathematics, and code generation. It utilizes blocksparse attention and LongRope technology to support contexts up to 128K tokens.
Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
32
Key-Value Heads
8
Position Embedding
ROPE
RoPE Theta
1,000,000
Activation Function
Gated GELU
Dimensions
Base Precision
FP16
Hidden Dimension Size
4,096
Number of Layers
32
FFN Intermediate Size (Dense)
14,336
Tokenizer
Vocabulary Size
100,352
About Phi-3
Microsoft's Phi-3 models are small language models designed for efficient operation on resource-constrained devices. They utilize a transformer decoder architecture and are trained on extensively filtered, high-quality data, including synthetic compositions. This approach enables a compact yet capable model family.