Falcon H1

Falcon-H1 is the Technology Innovation Institute's (Abu Dhabi) hybrid language-model family combining Transformer attention with Mamba-2 state-space layers in a parallel design with tunable ratios — attention's expressiveness with SSMs' linear scaling and lower memory footprint. It spans 0.5B to 34B parameters (plus 1.5B-Deep and a 7B reasoning variant, Falcon-H1R, hitting ~88% on AIME 24), supports 18 languages natively, and TII reports the 34B model matching 70B-class transformers while the 0.5B rivals 2024-era 7B models. Instruct variants are available on AWS Bedrock Marketplace and SageMaker JumpStart.\n\nLimitations: it ships under TII's custom Falcon License rather than a standard OSI license, so commercial terms need review. The hybrid architecture is less widely supported by inference tooling than pure transformers. As a mid-2025 release it trails late-2025 frontier models on absolute capability, though it wins on efficiency per parameter.

TII-reported: Falcon-H1R-7B ~88.1% AIME 24, 83.1% AIME 25