Overview

Falcon-H1R 7B was trained from Falcon-H1-7B-Base with supervised fine-tuning on long reasoning traces, then reinforcement learning with GRPO. Its hybrid design mixes attention with Mamba2 state-space layers, which keeps long reasoning chains cheaper to generate.

Highlights

  • About 7.6B parameters
  • Hybrid Transformer and Mamba2 architecture
  • Trained for long chain-of-thought reasoning
  • Supported in Transformers, vLLM and SGLang

Typical uses

Math and reasoning tasks, and research on efficient test-time scaling.

Things to consider

  • Released under the Falcon LLM License, not a standard open-source license.
  • A small model; knowledge-heavy tasks favor larger models.

Specifications are taken from the model card on Hugging Face and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.