Intern-S2-Mobius splits memory from reasoning, gets 4x speedup

Intern-S2-Mobius, an experimental transformer, runs inference nearly 4x faster than a same-accuracy baseline, per an arXiv preprint. A 7B version of the same architecture, trained from scratch, matches a standard 7B transformer's downstream scores using only 62.6% of the training data. For a lab, that's fewer GPU-hours to train and fewer to serve, if the results hold outside the authors' own tests.

The architecture, called Mobius-v0, replaces the usual blend of memory and reasoning inside a transformer with two separate parts. One shared Memory module, built from the feed-forward network, holds knowledge vectors. Multiple Reasoner modules, built from self-attention, query that Memory repeatedly, using hidden states as both a cache and a carrier between queries.

Two versions test the claim. The 7B model, trained from scratch, hits the same downstream scores as a standard 7B transformer on 62.6% of the training data. Intern-S2-Mobius, continually pretrained from Qwen3.5-35B, matches its baseline's accuracy while running nearly 4x faster.

The preprint doesn't say what it costs to retrofit a model beyond the Qwen3.5 case, and it doesn't compare Mobius against other efficiency techniques. Both results come from the authors' own benchmarks, not an independent replication.