Abu Dhabi’s Technology Innovation Institute has released Falcon H1R 7B as an open-weight reasoning model.
The Technology Innovation Institute (TII), Abu Dhabi has released Falcon H1R 7B, a 7-billion-parameter, reasoning-focused large language model, positioning it as a major open-weight challenge to much larger proprietary and open models. The full model weights and code are publicly available on Hugging Face, accompanied by a live inference demo via Falcon Chat and a detailed technical report outlining architecture, training, and evaluation.
TII claims Falcon H1R 7B can out-reason models up to seven times its size, outperforming Alibaba Qwen 32B and 47B and larger variants of Nvidia Nemotron, calling into question the assumption that better reasoning demands ever-larger parameter counts.
At the core of the model is a hybrid Mamba–Transformer architecture, combining Transformer attention with Mamba state-space model (SSM) layers. This design enables linear-time sequence processing, lower memory usage, and reduced compute cost for long chain-of-thought reasoning. According to TII, Falcon H1R 7B processes around 1,500 tokens per second per GPU at batch size 64, nearly twice the throughput of Qwen3 8B, while sustaining performance on long responses.
On the AIME 2025 mathematical reasoning benchmark, Falcon H1R 7B scored 83.1%, beating larger open models such as Apriel-v1.6-Thinker (15B) and OLMo 3 Think (32B), and approaching proprietary systems including Claude 4.5 Sonnet and Amazon Nova 2.0 Lite. It also delivered 68.6% on the LCB v6 coding benchmark, the highest score claimed among tested models.
Released under the Falcon LLM License 1.0, the model allows royalty-free commercial use with mandatory attribution, reinforcing TII’s push to make efficient, high-density reasoning a credible open-weight alternative to closed frontier models.
