A new model architecture separates what an AI knows from how it reasons, and the split makes inference dramatically cheaper.
Researchers describe Mobius-v0, an architecture built around a single shared memory module made of feed-forward network layers that stores knowledge as vectors, paired with multiple attention-based reasoner components that repeatedly query that memory while working through a problem. Instead of baking knowledge into every layer of a standard transformer, Mobius keeps the knowledge store separate and lets reasoning modules pull from it as needed, passing hidden states back and forth as a kind of working cache. A 7B-parameter model trained from scratch this way matched a standard 7B transformer's benchmark scores while using only 62.6% of the training data. A second model, Intern-S2-Mobius, was built by continuing to train Qwen3.5-35B on this architecture, and it matched baseline accuracy while running nearly four times faster end to end.
Training data and inference speed are the two costs that determine whether a model is actually usable at scale, and this paper claims wins on both without sacrificing accuracy. If decoupling memory from reasoning holds up beyond this one paper's benchmarks, it suggests a chunk of what transformers spend compute on is redundant knowledge storage duplicated across layers, a structural inefficiency rather than a data problem.
That is a big claim from a single arXiv preprint with no independent replication yet, so file the 4x speedup under promising, not proven.