Yi-9B
Parameters
9B
Context Length
4K
Modality
Text
Architecture
Dense
License
Apache 2.0
Release Date
6 Mar 2024
Knowledge Cutoff
Jun 2023
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
4,096 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
No evaluation benchmarks for Yi-9B available.
Rankings
Overall Rank
-
Coding Rank
-
About Yi-9B
The Yi-9B model is a sophisticated dense transformer-based large language model developed by 01.AI, designed to optimize the trade-off between parameter count and reasoning depth. It serves as a performance-oriented extension of the foundational Yi-6B model, engineered through a process of architectural expansion and multi-stage incremental training. By increasing the model's depth and continuing pre-training on an additional 0.8 trillion high-quality tokens, the developers have produced a model that excels in technical domains such as mathematics and code generation while maintaining robust bilingual fluency in English and Chinese.
Technically, Yi-9B utilizes a decoder-only architecture that mirrors the established Llama framework, enabling immediate compatibility with the broader ecosystem of LLM tools and libraries. Key architectural features include Grouped-Query Attention (GQA) to improve inference throughput and reduce memory overhead, and SwiGLU activation functions within the feed-forward layers for enhanced representational capacity. The model employs Rotary Position Embedding (RoPE) to manage sequence data and utilizes Root Mean Square Layer Normalization (RMSNorm) to stabilize training dynamics across its 44 layers.
Designed for computational efficiency, Yi-9B is particularly suited for deployment in resource-constrained environments, including consumer-grade hardware. Its extensive training on a total of 3.9 trillion tokens provides the model with a strong knowledge base for complex reasoning, reading comprehension, and common-sense logic. This makes it an effective choice for developers building AI-native applications that require a balance of high-performance technical reasoning and efficient local execution.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
32
Key-Value Heads
4
Attention Head Dimension
128
Position Embedding
Absolute Position Embedding
RoPE Theta
10,000
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
4,096
Number of Layers
44
FFN Intermediate Size (Dense)
11,008
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
64,000