Falcon-H1-34B

Falcon-H1-34B is the largest model in TII Abu Dhabi's Falcon-H1 series, released on May 21, 2025. Each block combines attention heads and Mamba2 state-space heads in parallel, a hybrid design aimed at faster inference, lower memory use, and stronger long-range generalization than pure transformers.

The family spans 0.5B to 34B parameters with native training in 18 core languages scalable to over 100. TII reports the series outperforming models twice its size on reasoning, math, and coding, and the weights are integrated with vLLM, llama.cpp, SGLang, and MLX. Note the series ships under TII's custom Falcon license rather than a standard OSI license.