The Technology Innovation Institute (TII) of Abu Dhabi released Falcon-H1R 7B on January 5, a 7-billion-parameter reasoning model that outperforms rivals up to seven times its size. The model scored 88.1% on the AIME-24 math benchmark—beating ServiceNow Apriel 1.5 (15B) at 86.2%—and achieved 68.6% on LiveCodeBench v6 coding tasks.
Falcon H1R combines three design innovations: a hybrid Transformer-Mamba2 backbone, 256k token context window, and training that mixes supervised long-form reasoning with reinforcement learning via GRPO. The architecture enables up to 1,500 tokens/sec/GPU at batch 64, nearly double the speed of comparable models like Qwen3-8B.
TII released the model as open-source under the Falcon TII License. Developers can access it via Hugging Face with full technical documentation. The release challenges the assumption that AI reasoning requires massive parameter counts.
