Tasks and inputs
Sage-SciPerception is listed for Feature extraction, Multimodal. 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
Sage-SciPerception is Shanghai AI Lab's set of projectors feeding crystals, spectra, biological sequences and biosignals into a language model's input.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open download; licence not yet specified, so treat as research-only. No hosted price.
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
Sage-SciPerception is not a language model. It is a growing set of small trained projectors from Shanghai AI Laboratory that translate scientific data into the embedding space a language model reads. A crystal structure, an infrared spectrum, a protein sequence or an ECG trace goes through its usual scientific encoder, the projector maps those features into token embeddings, and the language model answers in natural language or a structured format.
The release appeared on 6 November 2026 under the Chinese name 书生·元征.
Each projector is trained for one task and pinned to a specific language model revision recorded in its config.json. Publisher-reported results:
| Task | Metric | Score |
|---|---|---|
| JARVIS-DFT formation energy | MAD/MAE (higher is better) | 15.03 |
| MPBind protein interface sites | per-protein F1 | 61.12% |
| QM9S structure from spectra | Acc@1 | 49.75% |
| GUE splice-site classification | MCC | 87.87% |
| BEACON ncRNA family | accuracy | 97.35% |
| SimXRD crystal system | accuracy | 63.60% |
| SciTS ECG classification | macro-F1 | 71.53% |
| SciTS sleep staging | macro-F1 | 72.01% |
| ATOM3D protein structure ranking | Spearman | 35.36% |
Read the formation-energy row carefully: the metric is a ratio where higher is better, not an error.
The projected vectors are spliced into the language model's fully templated prompt embeddings, including system and assistant prefixes, and sent to vLLM through prompt_embeds on the /v1/completions endpoint. Start vLLM with --enable-prompt-embeds. Feature extraction and projection happen on the client; the server only generates.
That design has consequences:
[num_atoms, 64].The licence has not yet been specified for this release. Until it is, treat it as research-only and do not build commercial products on it.
If your group already runs one of these encoders, the projector route lets a chat model reason about your data without fine-tuning the whole model. Pick one task you have labelled data for, run the projector against the paired language model, and compare its answers with your existing specialist model, not just with a text-only prompt. Shanghai AI Lab's related language models include Sage S2.1 and Intern-S2-397B; each projector's config.json names the exact model it needs. Other scientific and multimodal models are in the catalogue.
Compare the model on the work you need to complete.
Sage-SciPerception is listed for Feature extraction, Multimodal. 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.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
NVIDIA
Alpamayo 2 Edge is NVIDIA's open 2B-class driving VLA that outputs trajectories with causal explanations, sized for DRIVE AGX Thor in-vehicle use.
inclusionAI
Ling 3.2 Flash is inclusionAI's 560B-A25B multimodal MoE with a 1M-token window, built for visual agents, 3D tasks and medical image reasoning.
Swiss AI
Apertus 1.5 8B is the Swiss AI Initiative's open multilingual model, now taking text, images and audio with a 262K-token context window.
kyutai
REKI-2B is Kyutai's open 2.4B robot vision-language-action model with cheap register-token memory, pretrained only on public data.
black-forest-labs
FLUX 3 Action is Black Forest Labs' open 7B world action model for robots, predicting action chunks and future video from cameras, state and text.