Falcon3-3B
Parameters
3B
Context Length
33K
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
32,768 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
No evaluation benchmarks for Falcon3-3B available.
Rankings
Overall Rank
-
Coding Rank
-
About Falcon3-3B
The Falcon3-3B model is part of the Falcon 3 family of open foundation models developed by the Technology Innovation Institute (TII). This model is designed for a balance of performance and efficiency, enabling its deployment on a range of computing infrastructures, including smaller devices. It is developed to support advancements in capabilities related to science, mathematics, and code generation. The Falcon 3 series includes both base models for general-purpose generative tasks and instruct models for conversational applications, emphasizing accessibility in advanced artificial intelligence systems.
Architecturally, Falcon3-3B employs a transformer-based causal decoder-only design. It incorporates 22 decoder blocks, contributing to its processing depth. For attention mechanisms, the model utilizes Grouped Query Attention (GQA) with 12 query heads and 4 key-value heads, along with a wider head dimension of 256. This configuration supports efficient inference operations. The model integrates SwiGLU as its activation function and RMSNorm for normalization, in addition to using Rotary Position Embeddings (RoPE) with a high value to handle extended context. It also leverages Flash Attention 2 for optimized memory and speed during operations.
The Falcon3-3B model, particularly its instruct variant, supports a context length of up to 32,768 tokens, while the base version supports 8,192 tokens. It is engineered to perform on tasks such as reasoning, language understanding, instruction following, and mathematical problem-solving. The model has been trained to support four languages: English, French, Spanish, and Portuguese. Its design considerations include the availability of quantized versions, such as int4, int8, and 1.58 Bitnet, which further enhance its efficiency and suitability for resource-constrained environments.
Technical Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
24
Key-Value Heads
6
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
1,536
Number of Layers
28
FFN Intermediate Size (Dense)
9,216
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.