Falcon H1R 7B
Technology Innovation Institute · 2026-01-05 · 7.6B parameters
Falcon H1R 7B is TII's 7B hybrid reasoning model released in January 2026, combining a Transformer attention backbone with Mamba2 state-space layers for memory-efficient long-context processing up to 256K tokens. It is trained with supervised long-form reasoning traces and reinforcement learning via GRPO, targeting math, coding, and logic tasks, and achieves competitive performance against reasoning models several times its parameter count.