Shanghai AI Laboratory released Intern-MemDec-4B, a plug-in memory decoder for Intern-S2 biology models
Shanghai AI Laboratory released Intern-MemDec-4B on Hugging Face, a 4B-parameter memory decoder that attaches to Intern-S2 backbones to add biology domain knowledge without retraining. It improved Biology-Instructions average score from 56.92 to 60.32 when paired with Intern-S2-Preview-397B. The model is not standalone and requires a compatible Intern-S2 backbone and router.
China context
- Original name
- 上海人工智能实验室
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the weights from Hugging Face and integrate the memory decoder with Intern-S2 backbones, but must source the router/fusion configuration separately.
- For investors
- The modular memory approach could reduce the cost of domain-specific model updates for biotech companies, shifting spending from full fine-tuning to plug-in modules.
Translated from Chinese. Quotes and facts link to the original sources.
Intern-MemDec-4B is a modular domain-extension component for Intern-S2 foundation models. It processes the same context as the backbone in parallel, and a lightweight token-level router dynamically combines their predictions. The memory is trained by compressing retrieval-based evidence from biology data into a parametric module, so inference does not need the original datastore. Evaluation on 21 Biology-Instructions tasks with Intern-S2-Preview-397B shows an average score increase from 56.92 to 60.32, with gains on DNA, RNA, and protein tasks while general capabilities remain broadly comparable.
The architecture uses a parallel decoding scheme where backbone and memory decoder produce next-token predictions independently, and a token-level router fuses them. This avoids full-model fine-tuning and allows domain knowledge to be added as a plug-in. The memory is trained via retrieval compression, so no external datastore is needed at inference. The release covers biology only, with task families spanning DNA, RNA, proteins, and biomolecular interactions.
Biotech and pharma teams using Intern-S2 can now add specialized biology knowledge without retraining the 397B backbone, reducing compute cost and preserving general reasoning. Competitors offering domain-adapted scientific models face a modular alternative that lowers the barrier to updating models as biological knowledge evolves.
For enterprises in drug discovery, genomics, or protein engineering, Intern-MemDec-4B offers a way to extend a general scientific model with domain expertise without the cost and risk of full fine-tuning. It may reduce infrastructure needs since no retrieval datastore is required at inference.
Watch for release of compatible router/fusion configurations and documentation for attaching Intern-MemDec-4B to Intern-S2 backbones. The paper (MemSFT) may detail training methodology and cross-domain generalization. Adoption can be checked via Hugging Face downloads and community forks.