Mixtral-8x7B-v0.1
Active Parameters
46.7B
Context Length
33K
Modality
Text
Architecture
Mixture of Experts (MoE)
License
Apache 2.0
Release Date
9 Dec 2023
Knowledge Cutoff
Nov 2022
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
5x RTX 4090
24GB VRAM
Datacenter
2x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
32,768 tokens
Consumer
5x RTX 4090
24GB VRAM
Datacenter
2x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#144
| Benchmark | Score | Rank |
|---|---|---|
General Text | 1197 | 148 |
Rankings
Overall Rank
#144
Coding Rank
-
About Mixtral-8x7B-v0.1
Mixtral-8x7B-v0.1 is a generative large language model developed by Mistral AI, distinguished by its Sparse Mixture of Experts (SMoE) architecture. This design enables the model to process information efficiently by conditionally activating a subset of its parameters for each input. Its primary purpose is to facilitate advanced text generation and comprehensive language understanding across a diverse range of applications.
The model is built upon a decoder-only transformer architecture. It integrates a Mixture-of-Experts layer where each layer comprises eight distinct feedforward blocks, known as 'experts'. A router network dynamically selects two of these experts to process each token, subsequently combining their outputs additively. This mechanism permits the model to leverage a substantial total parameter count of 46.7 billion while maintaining a significantly lower active parameter count of 12.9 billion per token during inference, thereby optimizing the balance between model capacity and computational efficiency. The architecture further incorporates Grouped Query Attention (GQA) and supports Flash Attention for enhanced performance.
Mixtral-8x7B-v0.1 supports a context length of 32,000 tokens, allowing it to process and generate responses based on extensive textual inputs. The model demonstrates proficiency in multilingual tasks, supporting English, French, Italian, German, and Spanish. It also exhibits strong performance in code generation tasks. The model can be fine-tuned for instruction-following tasks, making it a suitable foundation for building interactive applications that require precise adherence to user commands.
Technical Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
32
Key-Value Heads
8
Attention Head Dimension
128
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
Swish
Dimensions
Hidden Dimension Size
4,096
Number of Layers
32
FFN Intermediate Size (Dense)
14,336
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
32,000
Mixture of Experts
Total Expert Parameters
7.0B
Number of Experts
8
Active Experts
2
Shared Experts
-
FFN Intermediate Size (per Expert)
14,336
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.