Mistral-7B-Instruct-v0.1
Parameters
7.3B
Context Length
8K
Modality
Text
Architecture
Dense
License
Apache 2.0
Release Date
27 Sept 2023
Knowledge Cutoff
-
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
8,192 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
Rank
#161
| Benchmark | Score | Rank |
|---|---|---|
General Text | 1110 | 152 |
Rankings
Overall Rank
#161
Coding Rank
-
About Mistral-7B-Instruct-v0.1
The Mistral-7B-Instruct-v0.1 model is an instruction-tuned variant of the Mistral-7B-v0.1 generative text model, developed by Mistral AI. Its primary purpose is to facilitate conversational AI and assistant tasks by precisely interpreting and responding to instructional prompts. This model is designed for efficiency, providing a compact yet performant solution for language processing applications.
Architecturally, Mistral-7B-Instruct-v0.1 is a decoder-only transformer model. It incorporates several advancements to enhance computational efficiency and context management. These include Grouped-Query Attention (GQA) for accelerated inference and Sliding-Window Attention (SWA), which enables processing of longer input sequences more effectively by attending to a fixed window of prior hidden states. The model utilizes Rotary Position Embedding (RoPE) for positional encoding and employs RMS Normalization. Its tokenization is handled by a Byte-fallback BPE tokenizer.
Regarding its capabilities, Mistral-7B-Instruct-v0.1 is applicable across various text-based scenarios. It is adept at generating coherent text, answering questions, and performing general natural language processing tasks. Specific applications include conversational AI systems, educational tools, customer support interfaces, and knowledge retrieval agents. Its design also supports real-time content generation and energy-efficient AI deployments due to its optimized architecture.
Technical Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
32
Key-Value Heads
8
Attention Head Dimension
-
Position Embedding
ROPE
RoPE Theta
10,000
Sliding Window Attention
Yes
Sliding Window Size
4,096
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
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
About Mistral 7B
Mistral 7B, a 7.3 billion parameter model, utilizes a decoder-only transformer architecture. It features Sliding Window Attention and Grouped Query Attention for efficient long sequence processing. A Rolling Buffer Cache optimizes memory use, contributing to its design for efficient language processing.