Mixtral-8x22B-v0.1
Active Parameters
176B
Context Length
66K
Modality
Text
Architecture
Mixture of Experts (MoE)
License
Apache 2.0
Release Date
10 Apr 2024
Knowledge Cutoff
-
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
19x RTX 4090
24GB VRAM
Datacenter
6x NVIDIA A100
80GB VRAM
Apple Silicon
4x Apple M3 Max
128GB VRAM
65,536 tokens
Consumer
20x RTX 4090
24GB VRAM
Datacenter
6x NVIDIA A100
80GB VRAM
Apple Silicon
4x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#150
| Benchmark | Score | Rank |
|---|---|---|
Summarization ProLLM Summarization | 0.587 | 26 |
General Text Text Arena | 1228 | 95 |
Rankings
Overall Rank
#150
Coding Rank
-
About Mixtral-8x22B-v0.1
Mixtral-8x22B-v0.1 is a large language model developed by Mistral AI, characterized by its Sparse Mixture-of-Experts (SMoE) architecture. This design approach enables the model to handle a wide array of natural language processing tasks efficiently, including text generation and comprehension. The model's architecture is engineered to balance computational demands with high performance, making it suitable for applications requiring substantial language understanding capabilities.
The core of Mixtral-8x22B-v0.1's architecture involves a system of eight specialized neural network experts, each contributing to the model's overall processing capacity. While the model comprises a total of 176 billion parameters, its sparse activation mechanism ensures that only two of these experts are actively engaged for any given input token. This selective activation results in an active parameter count of approximately 39 billion, significantly reducing the computational load during inference compared to a densely activated model of equivalent total size. The model operates with a decoder-only transformer framework and utilizes sparse activation patterns for optimized performance.
Mixtral-8x22B-v0.1 demonstrates proficiency across multiple domains, including multilingual understanding, mathematical problem-solving, and code generation. It is fluent in languages such as English, French, Italian, German, and Spanish. Furthermore, it incorporates native function calling capabilities, enhancing its utility in integrated application environments. These characteristics make it a robust tool for diverse use cases such as chatbot development, content creation, document summarization, and complex question-answering systems that benefit from its ability to process extensive context windows.
Technical Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
48
Key-Value Heads
8
Attention Head Dimension
-
Position Embedding
ROPE
RoPE Theta
1,000,000
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
1,024
Number of Layers
56
FFN Intermediate Size (Dense)
16,384
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
32,000
Mixture of Experts
Total Expert Parameters
22.0B
Number of Experts
8
Active Experts
2
Shared Experts
-
FFN Intermediate Size (per Expert)
16,384
Dense Layers Before MoE
-
About Mixtral
The Mixtral model family, developed by Mistral AI, employs a sparse Mixture-of-Experts (SMoE) architecture. This design utilizes multiple expert networks per layer, where a router selects a subset to process each token. This enables large total parameter counts while maintaining computational efficiency by activating only a fraction of parameters per forward pass.