Falcon H1R 7B

- Advanced reasoning capabilities in mathematics, coding, logic, and instruction-following in a highly efficient 7B model.

- Combines an attention-based and state-space architectures for efficient sequence modeling under large batch sizes and long response lengths.

- Designed for low latency, reduced energy use, and flexible deployment on large batch sizes.

- Open and research-driven, released under the Falcon LLM license.

Try our Chat

Advanced Reasoning at 7B Scale

We designed Falcon H1R 7B to be practical, not just impressive to demonstrate.

Falcon H1R 7B is a compact, high-performance large language model designed to deliver world-class reasoning capabilities within a 7-billion-parameter architecture. Built on the Falcon H1 7B foundation, H1R introduces specialized reasoning capabilities that significantly enhances logical accuracy, mathematical problem-solving, and coding task performance, all while maintaining high efficiency and speed under large-batch size.

By combining architectural innovation with specialized training techniques, we’ve achieved benchmark-leading results while remaining fast, scalable, and accessible for real-world deployment.

Falcon H1R 7B proves that advanced reasoning no longer requires large model sizes, thereby setting a new benchmark for intelligent and cost-effective deployable AI.

This is a leap forward in the reasoning capabilities of compact AI systems showing how scientific precision and scalable design can go hand in hand.

Read our technical blog

Average
Performance Benchmark

See Details

Beating the Benchmarks

Mathematical Reasoning

High performance on AIME-24, AIME-25, and the recently introduced AMO-Bench demonstrating strong mathematical problem-solving at a fraction of the model size.

Code and Agentic Tasks

Best in class performance among models under 8B parameters.

General Reasoning

Matched or closely approached the performance of models with up to twice the parameter count while using substantially fewer resources.

Efficiency

Achieved 2x tokens per second per GPU at batch 64 without sacrificing accuracy compared to leading 8B transformers.

Designed for Efficient Intelligence

Falcon H1R 7B is optimized for environments where compute efficiency, latency, and scalability matter. With just 7B parameters, this model rivals and surpasses significantly larger open-source models, providing advanced reasoning at a fraction of the computational cost.

Whether you need advanced reasoning for enterprise workflows, developer tools, or research applications, Falcon H1R 7B delivers powerful intelligence that is efficient, accessible, and ready to deploy.

Hybrid Architecture for Robust Performance

Falcon H1R 7B is built on a hybrid Transformer-Mamba architecture, combining the strengths of attention-based reasoning with efficient sequence modeling. This architectural approach enables faster inference and improved long-context handling while maintaining strong reasoning depth.

The result is a new performance frontier.

We’ve incorporated specialized test-time reasoning training to unlock latent intelligence, so Falcon H1R 7B can reason more accurately and effectively without increasing parameter count.

World-Class Reasoning in a Compact Model

Falcon H1R 7B pushes the limits of what compact models can achieve, delivering advanced reasoning performance in a highly efficient architecture.

Benchmark-Leading Reasoning

The reference model for efficient, high-quality reasoning, Falcon H1R 7B delivers standout results across multiple competitive benchmarks.

Built for Real-World Deployment

Faster inference and lower latency for when performance efficiency is critical.

Falcon H1R 7B

Advanced Reasoning at 7B Scale