Falcon-H1

Released by Abu Dhabi's Technology Innovation Institute in May 2025, Falcon-H1 is a family of hybrid language models combining Transformer attention with Mamba2 structured-state-space components. The design targets the classic trade-off in modern LLMs: Transformers are expressive but expensive at long context due to quadratic attention, while SSMs scale linearly but lag on nuanced understanding.

The family spans 0.5B to 34B parameters, covering resource-constrained deployments up to large-scale inference, with an emphasis on multilingual support and extended input sequences. A January 2026 follow-up, Falcon-H1 Arabic (3B/7B/34B), topped the Open Arabic LLM Leaderboard. Weights are openly downloadable from Hugging Face under the TII Falcon license; independent third-party verification of some performance claims remains thin, as with most vendor-reported benchmarks.