Tasks and inputs
Intern-MemDec-4B is listed for Language & text, Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 19, 2026
Intern-MemDec-4B is Shanghai AI Lab's biology memory decoder that plugs into Intern-S2 backbones to add DNA, RNA and protein knowledge without retraining.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open download; no hosted price. No licence is stated for the weights, so check the repository before commercial use.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Fine-tuning a big model on biology usually costs you some of its general ability. Intern-MemDec-4B is Shanghai AI Laboratory's way around that alignment tax. Instead of changing the Intern-S2 backbone, it bolts on a separate 4B-class "memory decoder" that knows biology. Both run on the same context in parallel, and a small token-level router decides, at each step, how much each one's prediction counts.
It is not a chat model. Load it on its own and you get nothing useful.
The memory was trained (the method is called MemSFT) to imitate a retriever searching biological data. In effect it memorises what retrieval would have fetched, so at inference time there is no datastore to host or query. On biology prompts the router leans on the memory; on general prompts it keeps the backbone in charge, which is how general reasoning, instruction following and multimodal behaviour are supposed to survive.
With Intern-S2-Preview-397B on all 21 Biology-Instructions tasks, the publisher reports the average rising from 56.92 to 60.32. The detail is more mixed than the average:
| Task | Backbone alone | + memory |
|---|---|---|
| Promoter-enhancer interaction | 22.46 | 38.47 |
| RNA CRISPR on-target | 6.61 | 17.18 |
| DNA tf-m | 56.96 | 67.09 |
| Protein thermostability | 58.44 | 53.97 |
| Antibody-antigen | 40.24 | 36.44 |
Gains are concentrated in DNA and RNA tasks; several protein and interaction tasks get slightly worse. If your work is protein engineering, test before assuming it helps.
The underlying paper also shows a trained memory being reused across backbones of different sizes, which is the long-term appeal: swap domain memories in and out instead of maintaining fine-tuned forks.
Split your test set by subdomain (DNA, RNA, protein, interactions) and score backbone-only against backbone-plus-memory separately for each. Also rerun a handful of non-biology prompts to confirm general behaviour holds. Related Intern-S2 models in the catalogue include Intern-S2-397B and Intern-S2-Mobius; confirm compatibility against the memory's router configuration before pairing. For a domain model trained end to end, compare with GPT-Rosalind 6. More are listed under language and text.
Compare the model on the work you need to complete.
Intern-MemDec-4B is listed for Language & text, Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under the publisher’s model licence. Check runtime support and size the complete model and working context for your expected concurrency.