Yi-34B
Parameters
34B
Context Length
4K
Modality
Text
Architecture
Dense
License
Apache 2.0
Release Date
2 Nov 2023
Knowledge Cutoff
Jun 2023
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
4x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
4,096 tokens
Consumer
4x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#154
| Benchmark | Score | Rank |
|---|---|---|
Web Development WebDev Arena | 1183 | 101 |
General Text Text Arena | 1183 | 102 |
Rankings
Overall Rank
#154
Coding Rank
#119
About Yi-34B
The Yi-34B model, developed by 01.AI, is a 34-billion parameter large language model trained from scratch on a 3-trillion token multilingual corpus. This foundational model demonstrates strong capabilities in language understanding, commonsense reasoning, and reading comprehension. It is specifically engineered to support both English and Chinese languages, offering robust bilingual proficiency across various tasks. The model's design focuses on achieving a balance between high performance and efficient inference, making it suitable for a range of computational environments.
Architecturally, Yi-34B is built upon a modified decoder-only Transformer framework, drawing inspiration from the LLaMA implementation without being a direct derivative. A key technical feature is the incorporation of Grouped-Query Attention (GQA), which contributes to reduced training and inference costs compared to traditional Multi-Head Attention while maintaining performance. The model utilizes the SwiGLU activation function and RMS Normalization layers. Positional encoding is handled through a Rotary Position Embedding (RoPE) mechanism. These architectural choices aim to optimize model stability, convergence, and compatibility within the AI ecosystem.
Yi-34B is applicable to tasks requiring extensive language processing, such as long-form document summarization, detailed legal and technical document analysis, and complex multilingual question-answering systems. It also excels in the generation of multilingual content and instruction following. The base model supports a context length of 4,096 tokens, with specialized variants like Yi-34B-200K extending this capacity to 200,000 tokens, enabling processing of exceptionally long text sequences. Its design considerations allow for deployment on various hardware configurations, including consumer-grade GPUs, especially when employing quantization techniques.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
56
Key-Value Heads
8
Attention Head Dimension
-
Position Embedding
Absolute Position Embedding
RoPE Theta
5,000,000
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
7,168
Number of Layers
60
FFN Intermediate Size (Dense)
20,480
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
64,000