Fully open ARLoop: 1.2B looped transformer from scratch beats Qwen3.5-2B Base
hhsun1 · x · 2026-11-10
ARLoop is a family of fully open, pretrained-from-scratch looped transformers built with Attention Residuals. Its strongest model, ARLoop-1.2B, outperforms both Huginn-3.5B (a prior fully open looped LLM) and Qwen3.5-2B Base on benchmarks spanning commonsense reasoning, math, and coding.
The team releases the complete recipe: data preparation, training, intermediate checkpoints, and evaluation code — making it one of the best open resources for looped transformer/LLM research.
Related event: ARLoop: Fully Open-Source Looped Transformers, 1.2B Beats Huginn-3.5B(5 posts)→
More from Models
- Google engineer spotlights AI Studio redesign, name-drops unreleased Gemini 4 Carbon — ammaar · 2026-11-12
- All frontier models cheat on HoneyBench v0.2, with cheat rates from 12% to 86% — ChowdhuryNeil · 2026-11-12
- Google DeepMind Code Summit talk urges devs to prepare for massive inference speed gains — altryne · 2026-11-12
- Never letting the model write directly: custom tools make Fable 5.2 prose feel human — Flomerboy · 2026-11-12
- Vals AI analyzed 200K benchmark runs: 14 models swear in reasoning, Xiaomi MiMo V2.6 Pro is foulest — ziqiao_ma · 2026-11-12
- Decido-1: open 26B MoE decision model beats Jev 1.13, runs CPU-only on laptops — gauravapiscean · 2026-11-12