Intern-S2-Mobius separates memory and reasoning to speed up inference

The Mobius-v0 architecture stores knowledge in a shared memory and entrusts reasoning to iterative modules, with a gain of about four times in inference speed.

A team led by Kai Chen, with 46 co-authors, proposes to dissociate two functions usually mixed in a transformer: knowledge storage and reasoning. The Mobius-v0 architecture entrusts the first to a shared global memory, built on the feed-forward layers, and the second to several "reasoners" based on self-attention, applied iteratively to compose the steps.

Two results are highlighted. A 7 billion parameter model driven from zero achieves performance comparable to a reference transformer using only 62.6% of the drive data. And Intern-S2-Mobius, obtained by continuing Qwen3.5-35B pre-training, offers an acceleration of end-to-end inference close to a factor four, with equivalent performance on downstream tasks.

The work was deposited on arXiv on August 14, 2026.