Falcon3-7B
Parameters
7B
Context Length
32K
Modality
Text
Architecture
Dense
License
TII Falcon-LLM License 2.0
Release Date
17 Dec 2024
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
32,000 tokens
Consumer
2x 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 Falcon3-7B available.
Rankings
Overall Rank
-
Coding Rank
-
About Falcon3-7B
Falcon 3-7B is a state-of-the-art instruction-tuned language model developed by the Technology Innovation Institute (TII). This model variant is a component of the Falcon 3 family, which focuses on enhancing capabilities in scientific domains, mathematics, and code generation. It is engineered for efficiency and scalability, enabling deployment on a range of infrastructures, including those with limited computational resources. The model supports multilingual applications, with training encompassing English, French, Spanish, and Portuguese, and is designed to handle long-context tasks.
The architectural foundation of Falcon 3-7B is a transformer-based causal decoder-only design, incorporating 28 decoder blocks. It utilizes Grouped Query Attention (GQA) to optimize inference speed and memory efficiency, configured with 12 query heads and 4 key-value heads, and a head dimension of 256. The model integrates Rotary Positional Embedding (RoPE) with a high value of 1000042 to facilitate effective understanding and processing of extended contexts up to 32,000 tokens. Activation functions are implemented using SwiGLU, complemented by RMSNorm for normalization, contributing to training stability and efficiency. It is also optimized to utilize FlashAttention-3.
Falcon 3-7B was pretrained on a dataset comprising 14 teratokens of diverse web, code, scientific, and high-quality multilingual data. Following pretraining, it underwent further fine-tuning on 1.2 million samples, specifically tailored for STEM content, conversational interactions, code, and safety compliance. This comprehensive training regimen positions the model for robust performance across various applications, including scientific and mathematical problem-solving, multilingual content generation, and processing long-form textual information. Its design supports instruction-following tasks, making it suitable for educational tools, research assistance, and the generation of technical documentation.
Technical Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
32
Key-Value Heads
8
Attention Head Dimension
256
Position Embedding
ROPE
RoPE Theta
1,000,042
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
36
FFN Intermediate Size (Dense)
23,040
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
131,072
About Falcon 3
The TII Falcon 3 model family comprises open-source, decoder-only language models (1B-10B parameters) designed for efficiency. Key innovations include an extended 32K token context window, Grouped-Query Attention (GQA), and specialized versions for scientific and code-oriented applications. Some variants integrate Mamba-based architectures.