Falcon3 7B Instruct
Falcon3 7B Instruct is the central pretrained scale in the Falcon3 transformer family. It supports English, French, Spanish and Portuguese with a 32K context.
How this Falcon3 size was made
TII reports 14 trillion pretraining tokens and 1.2 million post-training examples. The 28-block architecture uses grouped-query attention, SwiGLU and RMSNorm. Falcon Arabic later adapted Falcon3-7B through additional Arabic training.Falcon3-7B-Instruct model card
Keep the language variants distinct
Practical guidance. Use this checkpoint for its documented languages. If Arabic is required, compare the dedicated Arabic family directly. A shared base name does not mean the original tokenizer and a later adapted tokenizer represent Arabic equally well.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | December 2024 |
| Access | Downloadable |
| Model identifiers | tiiuae/Falcon3-7B-InstructFalcon3-7B-Base |
| Context window | 32,768 tokens Instruction checkpoint limit, corroborated by its max_position_embeddings configuration. Falcon3-7B-Instruct model card |
| License | TII Falcon-LLM License 2.0Falcon3-7B-Instruct model card |
| Weight access | Weights remain published.Falcon3-7B-Instruct model card |
| Variants | Base and Instruct; this page describes InstructFalcon3-7B-Instruct model card |
Related models
A 10B multilingual transformer language model created by extending the depth of Falcon3-7B-Base.
Launch date December 2024Downloadable