Key features
- Unified understanding and generation across 7 modalities through one
natural-language interface.
- Seven modalities, one 8B backbone (protein / RNA / DNA / molecule /
weather / medical-image / text) via a modality router.
- Native scientific encoders/decoders (ESM-2, RNA/DNA ConvFormers, molecular
graph encoder, Swin-ViT weather tower, SAM-based image path) preserve domain
structure a generic tokenizer would destroy.
Capabilities
Table with columns: Modality, Understanding, Generation| Modality | Understanding | Generation |
|---|
| Protein | ✅ | — |
| RNA | ✅ | ✅ |
| DNA | ✅ | — |
| Molecule | ✅ | ✅ |
| Weather | — | ✅ |
| Medical image | — | ✅ |
| Text | ✅ | ✅ |
Understanding = classification / regression / scientific QA. Generation: RNA sequence design · Molecule text → SMILES · Weather 10-day global ERA5 0.25° forecast · Medical-image text-prompted segmentation (SAM 3-based; Meta SAM License).
Benchmarks
神珍 (8B backbone, ~11B total) vs Biology-Instructions (Llama-3.1-8B,
text-token, no scientific encoders) and Intern-S1-Pro (~1T MoE scientific
model). Bold = best; underline = second-best.
Biological sequence understanding
Table with columns: Task, Metric, 神珍 (~11B), Biology-Instructions (8B), Intern-S1-Pro (~1T)| Task | Metric | 神珍 (~11B) | Biology-Instructions (8B) | Intern-S1-Pro (~1T) |
|---|
| DNA · Epigenetic marks (EMP) | MCC | 71.99 | 3.64 | 14.02 |
| DNA · Promoter det. 300bp (PD300) | MCC | 91.17 | 58.18 | 82.65 |
| DNA · Core-promoter (CPD) | MCC | 66.35 | 44.54 |
Aggregate over 20 biological-understanding benchmarks: 神珍 matches or beats the ~1T Intern-S1-Pro on 10/20 and the same-scale 8B text-token baseline on 16/20.
Molecule understanding (SMolInstruct)
Table with columns: Task, Metric, 神珍 (~11B), LlaSMol| Task | Metric | 神珍 (~11B) | LlaSMol |
|---|
| BBBP | Acc | 96.95 | 74.60 |
| HIV | Acc | 97.00 | 96.70 |
| SIDER | Acc | 71.00 | 70.70 |
| ClinTox | Acc | 92.36 | |
Earth-science forecasting — vs ECMWF HRES (day-10, global ERA5 0.25°)
Table with columns: Variable, Metric, 神珍 (~11B), ECMWF HRES (NWP)| Variable | Metric | 神珍 (~11B) | ECMWF HRES (NWP) |
|---|
| Z500 | RMSE ↓ | ≈680 | ≈800 |
| T2M | RMSE ↓ (K) | ≈2.5 | ≈2.9 |
| MSL | RMSE ↓ (Pa) | ≈625 | ≈740 |
神珍 tracks or beats the operational physics-based HRES system, with the advantage growing at longer lead times.
Medical-image segmentation
Mean Dice (%) on the BiomedParse test splits, 102,855 image–prompt pairs across
nine imaging modalities, versus six modality-native segmentation specialists.
Table with columns: Modality, # Samples, 神珍, BiomedParse, MedSAM, SAM, SAM3, DINO+MedSAM, DINO+SAM| Modality | # Samples | 神珍 | BiomedParse | MedSAM | SAM | SAM3 | DINO+MedSAM | DINO+SAM |
|---|
| All | 102,855 | 91.20 | 90.73 | 83.55 | 71.29 | 35.40 | 15.37 | 15.10 |
| CT | 45,306 |
Best overall Dice (All), and best on CT, MRI, pathology, dermoscopy, and endoscopy; on X-ray, Fundus, and Ultrasound the gap to BiomedParse is ≤ 0.5 Dice, and on the smallest split (OCT) it is 1.3.
Usage
Runs via the accompanying code repository (custom multimodal architecture).
git clone https://github.com/Shanghai-Academy-of-AI-For-Science/MKB && cd MKB
pip install -r requirements.txt # Python 3.10; transformers==5.0.0
hf download sais-org/MKB --local-dir ./model
export PYTHONPATH=$PWD/code
python code/inference.py --model_path model --greedy --max_new_tokens 64 \
--rna "GGATGCGATCATGTCTGCACTAACACACCGGATCCCATCAGAACTCCGAAGTTAAGCGTGCTTGGGCGGGAGTAGTACTAGGATGGGCGACCCCTTAGGAAGTACTCGTGTTGCATCCC" \
--system "You are a non-coding RNA family classifier. Output only the family name, no other text." \
--prompt $'<rna>\nWhich family does this non-coding RNA sequence belong to?'
All weights are contained in model.safetensors: the scientific
encoders/decoders (ESM-2, the Suiren molecular graph encoder, the RNA/DNA
ConvFormers, the Swin-ViT weather tower) and the fine-tuned SAM 3 branch used
for medical-image segmentation.
Each task has a specific --system prompt that fixes the output format; see
run_examples.sh in the repository for per-task examples, weather, and segmentation.
License
Composite license. 神珍's own components — the code, and all weights
except the SAM 3 branch — are Apache-2.0, built on Qwen3-VL (Apache-2.0) and
including merged ESM-2 (MIT) and Polaris/Suiren-derived encoders.
The medical-image segmentation branch embeds SAM 3 weights, which are
governed by Meta's SAM License (SAM_LICENSE.txt, shipped alongside these
weights). SAM 3 use is subject to that license, including its acceptable-use
restrictions (no military / weapons / illegal uses; Trade-Control compliance).
See THIRD_PARTY_LICENSES.md / NOTICE for the full third-party breakdown.
Citation
@misc{chen2026monkeykingbangunified,
title = {Monkey King Bang: A Unified Scientific Multimodal Foundation Model},
author = {Hesen Chen and Xinyu Su and Xiaomeng Yang and Yuetan Lin and Zixiong Yang and Junyi An and Fenglei Cao and Yifeng Jiao and Yunqi Zhang and Yuan Cheng and Zhiyu Tan and Hao Li and Libo Wu and Yuan Qi},
year = {2026},
eprint = {2607.20557},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2607.20557}
}