The TinyLlama project trains a compact 1.1B parameter language model following the Llama 2 architecture and tokenizer. Training uses 16 A100-40G GPUs over approximately 90 days to process 3 trillion tokens. The project releases intermediate checkpoints on HuggingFace, provides chat demos, includes speculative decoding examples with llama.cpp, and offers fine-tuning scripts for customizing the base model.