Falcon

Abu Dhabi's Technology Innovation Institute publishes Falcon as downloadable weights on Hugging Face, and the family now reaches far past its 2023 debut. Falcon 7B and 40B used Apache-2.0, while the gated 180B license bars offering it as a hosted service without TII approval. Falcon 3 followed with dense models from 1B to 10B plus a Mamba variant, and Falcon-H1 introduced a hybrid design that runs attention alongside Mamba2 state-space layers, in sizes from 0.5B to 34B covering 18 languages. Later additions include Falcon-H1R-7B, tuned with GRPO reinforcement learning to work through a problem before replying, Falcon-E models stored at 1.58 bits for edge devices, Falcon-H1-Tiny models near 90M parameters, and GPTQ, AWQ, and GGUF quantized builds. Recent checkpoints run in Transformers, vLLM, SGLang, or llama.cpp. Newer releases use the TII Falcon License, an Apache-2.0 derivative that is royalty free for commercial work but requires passing TII's acceptable use policy on to anyone you redistribute to. Pick Falcon for self-hosted chat, reasoning, or compact on-device models.