Researchers introduced Mobius-v0, an architecture that separates knowledge storage from reasoning by using a globally shared memory implemented as a feed-forward network alongside multiple self-attention Reasoner modules that iteratively query and retrieve relevant knowledge to perform compositional reasoning. A 7B model trained from scratch with this architecture matched baseline Transformer performance using significantly less training data, while the Intern-S2-Mobius variant, built on Qwen3.5-35B, achieved roughly 4x inference speedup with similar downstream performance.
