ARLoop releases fully open looped transformer recipe; 1.2B model beats Huginn-3.5B
Lianhuiq · x · 2026-11-11
The ARLoop team introduces a family of strong, fully open pretrained looped transformers with Attention Residuals, releasing the complete training recipe.
- Their strongest model, ARLoop-1.2B, outperforms both Huginn-3.5B (the prior best fully open looped LLM) and Qwen3.5-2B Base on commonsense reasoning, math, and coding benchmarks
- Everything is open-sourced: data preparation, training, intermediate checkpoints, and evaluation code
- The team notes that while looped LLMs are heating up and open weights are appearing, training recipes remain scarce — the gap this work fills
Related event: ARLoop: Fully Open-Source Looped Transformers, 1.2B Beats Huginn-3.5B(5 posts)→
More from Research
- Prof opens interactive NumPy-only ANN notebook to teach neural nets from scratch — GeostatsGuy · 2026-11-11
- Continual-ARC benchmark: leading agents fail to learn from experience, a fine-tuned 9B model doesn't — Med1_Ai · 2026-11-11
- Good conversations show 'pink noise' in semantic similarity, Science Advances finds — _akpiper · 2026-11-11
- Study: up to 50% of STEM PhD dissertations use AI writing, only 4.7% disclose it — paulnovosad · 2026-11-11
- AlphaGenome 2 commercial API enters Google Cloud private preview, weights promised later — anshulkundaje · 2026-11-11
- Stanford's DyCraft Uses Agentic Decomposition to Cut Physical Errors in Long-Horizon 4D Scene Generation — StanfordUniversity · 2026-11-11