[💻Github Repo](https://github.com/InternLM/Intern-S2-Mobius) • [🤗Model Collections](https://huggingface.co/collections/internlm/intern-s2) • [🌳Arch Space](https://github.com/InternLM/archspace)
## Introduction
We introduce **Intern-S2-Mobius**, a 35B foundation model built on the Mobius-v0 architecture realized by Xtuner and LMDeploy. Instead of binding knowledge storage and reasoning computation layer by layer as in conventional Transformer models, Mobius organizes knowledge into a globally shared **Memory** and lets multiple **Reasoners** iteratively query and refine hidden states against this shared repository.
This knowledge-reasoning separation gives Intern-S2-Mobius two native capabilities: **Backward Residual Connection**, where reasoning stages can access knowledge beyond their local layer hierarchy, and **Dynamic Latent Reasoning**, where deliberation, refinement, and multi-token prediction are internalized into high-density continuous states. Continual-pretrained from Qwen3.5-35B and further post-trained with SFT and RL, Intern-S2-Mobius preserves strong downstream capability while achieving substantially higher end-to-end inference efficiency, with nearly 4x speedup reported in the technical report.
### Features
- **Knowledge-reasoning decoupled architecture.** Intern-S2-Mobius separates knowledge vectors from reasoning operators by replacing layer-bound FFN knowledge storage with a globally shared Memory. This gives each Reasoner access to a broader knowledge space and improves knowledge compression compared with a standard Transformer layout.
- **Backward Residual Connection.** Through shared Memory, shallow and deep reasoning stages can access knowledge across the model rather than relying only on forward layer-wise information flow. This enables more flexible cross-layer knowledge composition and helps the model synthesize useful information in fewer reasoning steps.
- **Dynamic Latent Reasoning.** Mobius refines continuous hidden states through recurrent latent iteration before decoding. This internalizes part of the deliberation process, reduces reliance on long visible chain-of-thought, and dynamically allocates computation to different tokens.
- **Higher inference efficiency with concise reasoning.** On reasoning benchmarks, Intern-S2-Mobius reaches comparable or stronger scores than the Qwen3.5-35B baseline while producing markedly shorter reasoning traces and higher request throughput, leading to nearly 4x end-to-end inference speedup in the reported evaluation.
- **Strong general and scientific performance.** Intern-S2-Mobius improves the reported average score over Qwen3.5-35B on general reasoning benchmarks, and shows large gains on scientific tasks such as Biology-Instructions, Mol-Instructions, and MolecularIQ.
