A model needs to decipher the sequence code of DNA, understand the spatiotemporal evolution of meteorological fields, and identify lesions in medical images—tasks that used to require multiple specialized models. On July 20, the Shanghai Institute for Scientific Intelligence unveiled a systematic solution to this challenge: the open science multimodal foundation model Shenzhen (Monkey King Bang, MKB), which uses about 11 billion parameters to simultaneously handle six types of scientific data—DNA, RNA, proteins, small molecules, Earth systems, and medical images—within a single unified model, both understanding and directly producing results.

On the track of scientific intelligence, specialized models have already emerged for protein, molecular, meteorological, and medical imaging directions, each performing excellently. However, their underlying principles are not compatible: sequences rely on sequential dependencies, molecules emphasize atomic connection structures, meteorological fields focus on spatiotemporal evolution, and medical images emphasize local details. How to preserve these differences while enabling a single model to learn shared patterns has always been an unavoidable question for scientific foundational models. Shenzhen was created precisely to address this issue.

image.png

Its backbone is Qwen3-VL-8B as a shared main body, with separate data processing pathways for the six types of scientific data. The key lies in not simplifying all data into a single format, but preserving the sequential relationships of sequences, the structural connections of molecules, the spatial distribution of meteorological fields, and the local details of medical images before they enter the shared model. Proteins and nucleic acids still follow sequences, small molecules still follow structures, meteorological data retain spatial distribution, and medical images retain image details. These capabilities are integrated into a single model, and corresponding functions can be called according to tasks. In addition to answering text, Shenzhen can also directly generate RNA sequences, machine-readable SMILES molecular representations, global meteorological fields, and medical image segmentation results.

With about 11 billion parameters, Shenzhen has demonstrated competitive performance in tasks such as biological sequences, small molecules, meteorology, and medical imaging. In 20 tasks covering DNA, RNA, proteins, and different biological sequence relationship judgments, Shenzhen achieved the best results among three competing models in 9 tasks and ranked in the top two in 17 tasks. When compared individually, it outperformed the same-scale Biology-Instructions model in 16 tasks, while facing the 1 trillion parameter Intern-S1-Pro model, both models excelled in 10 tasks. This comparison illustrates one thing: parameter scale is not the only benchmark for evaluating scientific model capabilities; designing effective modeling methods tailored to different scientific objects is a more worthy direction to continuously refine.

Breaking away from biological sequences, Shenzhen also validated its performance in small molecules, meteorology, and medical imaging. In the small molecule domain, in six attribute understanding tasks on the public benchmark SMolInstruct, Shenzhen achieved the best or tied best results in four tasks. In meteorology, during offline evaluation, given an initial atmospheric field, the model made a forecast every 6 hours, continuing to simulate up to the 10th day. By the 10th day, metrics such as 500hPa geopotential height, 2-meter temperature, and sea level pressure reached or exceeded the level of operational numerical forecasts. In medical image segmentation, in an evaluation covering nine imaging modalities such as CT, MRI, and pathology, with over 100,000 image-text samples, the average Dice score of Shenzhen's predicted region compared to manually labeled regions reached 91.20, ranking first among seven evaluation methods. When categorized by image type, five categories ranked first, and the remaining four ranked second.

Shenzhen integrates four representative capabilities into a single unified model: understanding biological sequences, predicting small molecule properties or generating SMILES, rolling out global meteorological fields up to 10 days long, and completing medical image segmentation based on textual prompts. Different tasks call upon respective processing components, but share the same model backbone and jointly trained weights, eliminating the need for researchers to switch between multiple independent domain models.

For scientific models, open weights are just the first step. Whether it can provide executable code, examples, and clear input-output instructions directly determines whether the model can be verified and reused. This release includes model weights, inference code, example scripts, input instructions, and usage documentation, and completes the configurations needed for weather forecasting and medical image segmentation, supporting researchers to deploy locally or integrate into existing research workflows. Researchers can download the model or use the API through the Xinghe Qizhi Scientific Intelligence Open Platform, combining it with over 1,500 other scientific models and tools. They can also download the weights on Hugging Face and obtain inference code and examples on GitHub.

As the super brain of the system-level scientific intelligence agent "Great Sage" launched at the beginning of March this year, the name "Shenzhen" comes from the "Divine Treasure Iron" in the Journey to the West. The team hopes that this open model will carry diverse scientific tasks in a relatively concise form and encourage more researchers to participate in testing, using, and co-building it.