Falcon-LLM Team (2026) — ARXIV-2601.02346
Key takeaways
- A 7B reasoning-optimised model built from Falcon-H1-7B by further supervised fine-tuning and reinforcement-learning scaling; it matches or outperforms state-of-the-art reasoning models two to seven times its size on reasoning-heavy benchmarks.
- The authors attribute the result to data curation and targeted training rather than size, and frame efficiency in three dimensions at once: accuracy, tokens per solution, and inference speed, the last from the parallel hybrid architecture.
- Parallel test-time scaling with DeepConf — many samples, confidence- weighted — is where the inference speed pays off, and where the model is positioned as a backbone.
Standing in the anthology
The reasoning data mixture in LIT-119 is described as similar to this model’s, so it is the upstream of Falcon-H1-Tiny-R. Filed for that lineage and as an instance of the three-stage recipe (SOTA-129) that LIT-119 argues can be collapsed at tiny scale (SOTA-123).