Loopie Recurrent Transformer Outperforms Traditional Architectures at Equal Compute
Zitian Gao · hf · 2026-07-20
The Loopie series of recurrent Transformer models has been released, featuring two MoE versions: 20B total params (2B activated) and 6B total params (0.6B activated).
- Addressing Bottlenecks: Previously, increasing pre-training compute for recurrent Transformers was often less effective than simply scaling up parameters; Loopie overcomes this limitation.
- Performance: Ablation studies show that under an identical compute budget, its performance significantly exceeds traditional 30B-A3B baseline models.
- Reasoning Capabilities: Empowered by a novel post-training pipeline, the model achieves gold-medal-level performance in the 2025 IMO (International Mathematical Olympiad) and IPhO (International Physics Olympiad) without relying on external tools.
Related event: Loopie Cyclic Transformer Matches 30B Baselines with Fractional Tokens(7 posts)→
More from Models
- What are the best models to run on 48 GB of VRAM with two RTX 3090s? — ludos1978 · 2026-07-22
- Google releases Gemini 3.6 Flash as Gemini 3.5 Pro remains in testing — Ars Technica AI · 2026-07-22
- Google says Gemini 3.5 Pro is in partner testing as Gemini 4 pre-training starts — haider1 · 2026-07-22
- A benchmark chart puts a flash model around 5th place, but critics say it is far pricier — soumitrashukla9 · 2026-07-22
- Google introduces three new Gemini models focused on speed, token efficiency, and scale — Jason_perei · 2026-07-22
- How to Distinguish Genuine Token Efficiency from Shorter, Omissive Answers? — ruthstarkman · 2026-07-22