Shanghai Academy of AI for Science Opens ShenZhen, Challenging the Bigger-Is-Better Model

The Shanghai Academy of AI for Science has opened ShenZhen, an approximately 11-billion-parameter model built to handle six scientific data types. The Shanghai Academy of AI for Science ShenZhen multimodal foundation model covers DNA, RNA, proteins, small molecules, Earth systems, and medical images.

That scope creates the central tension. Scientific AI has usually advanced through specialized models designed around one domain, data structure, or prediction target. ShenZhen, also called Monkey King Bang or MKB, instead places domain-specific components around one shared language and vision backbone.

The academy says this smaller unified system matched or surpassed a roughly one-trillion-parameter scientific model on half of 20 biological understanding benchmarks. It also reported competitive molecule, weather, and medical imaging results. Those comparisons challenge the assumption that scientific AI must scale mainly through larger general-purpose models.

The release does not establish that one model can replace specialized scientific systems. Its benchmarks come from the development team, and many reported results remain unverified through independent studies. Still, the open weights and code give outside researchers a practical way to test the broader claim.

The ShenZhen Multimodal Foundation Model Connects Six Scientific Data Types

ShenZhen matters because it treats scientific data as distinct at the edges but connected at the model’s center.

The model uses Qwen3-VL-8B-Instruct as its shared backbone. Separate encoders and decoders preserve the structure of each scientific modality before information reaches that backbone. The complete system contains approximately 11 billion parameters, according to the academy.

This design avoids converting every input into ordinary text tokens. DNA and RNA require models that preserve sequence order. Small molecules need representations of atoms and chemical bonds. Weather data contains spatial fields that change over time, while medical images depend on fine local structures.

ShenZhen routes those different inputs through specialized components. Its model documentation identifies ESM-2 for proteins, convolutional transformer components for DNA and RNA, and a graph encoder for molecules. It also includes a Swin-ViT weather component and a SAM-based medical image path.

A modality router selects the relevant path for each task. The specialized component converts the input into a representation the shared backbone can process. A corresponding decoder then produces text, a scientific sequence, a spatial field, or an image mask.

That architecture supports both understanding and generation, although not equally across every modality. The model can classify or analyze protein, RNA, DNA, molecule, and text inputs. It can generate RNA sequences, molecular SMILES strings, weather forecasts, medical image segmentations, and text.

SMILES is a text notation for representing molecular structures in a computer-readable form. Image segmentation identifies the pixels belonging to a requested anatomical structure or abnormality. These outputs require different decoders even when the reasoning layer is shared.

The release therefore represents more than a science chatbot with several file readers. Its components produce native scientific outputs that researchers can pass into downstream software. A generated weather field or segmentation mask has different operational value from a prose description of the same data.

This approach also distinguishes ShenZhen from models that place scientific papers, charts, and diagrams inside a conventional vision-language training pipeline. Such systems can answer questions about scientific content without directly modeling DNA bases, molecular bonds, or atmospheric grids.

The academy has released model weights, inference code, example scripts, configuration files, and input guidance. Researchers can obtain the materials through Hugging Face and GitHub, while China’s NovaInspire platform offers additional access.

The model page lists an Apache 2.0 license for most components. However, the medical segmentation branch includes SAM-related weights governed by Meta’s separate license. Users must review those restrictions before deploying the complete package.

The release also connects ShenZhen with Dasheng, a scientific research agent introduced by the academy in March 2026. Dasheng coordinates models and tools through natural-language requests. ShenZhen now serves as a shared modeling layer for several scientific operations within that system.

That context explains why unification matters to the academy. A research agent becomes harder to operate when every task requires an unrelated model, environment, interface, and output format. A shared backbone can reduce that fragmentation without erasing each domain’s native structure.

An 11-Billion-Parameter Model Is Pressuring the Scale-First Approach

The most consequential benchmark is not a single score. It is ShenZhen’s reported performance against a model nearly 100 times larger.

The academy evaluated biological understanding across 20 tasks involving DNA, RNA, proteins, and relationships between biological sequences. It reported that ShenZhen placed first on nine tasks and ranked within the top two on 17.

Against Biology-Instructions, an eight-billion-parameter text-token baseline, ShenZhen reportedly scored higher on 16 of 20 tasks. That comparison supports the value of scientific encoders when the backbone sizes remain relatively close.

The more provocative comparison involves Intern-S1-Pro. Its developers describe it as a roughly one-trillion-parameter mixture-of-experts scientific model. A mixture-of-experts model routes each input through only part of its total network, reducing the computation required for each token.

ShenZhen and Intern-S1-Pro each led the other on ten biological tasks, according to the release. That result does not make the models equivalent. It does suggest that parameter count alone provides an incomplete measure of scientific capability.

Intern-S1-Pro follows the scale-first route more directly. Its predecessor, Intern-S1, used a 241-billion-parameter mixture-of-experts architecture with 28 billion active parameters. The team reported continued pretraining on five trillion multimodal tokens, including more than 2.5 trillion scientific tokens.

The current Intern-S1 repository presents Intern-S1-Pro as a trillion-scale model for scientific reasoning. Its scope includes scientific text and images, which makes it an important reference point. Its architecture and training strategy remain substantially different from ShenZhen’s domain-routed design.

Those differences matter when interpreting the biological comparison. ShenZhen uses encoders selected for scientific structures, while Intern-S1-Pro pursues wider scientific reasoning at much greater scale. A focused benchmark can favor the architecture more closely aligned with its input format.

ShenZhen’s strongest reported biological results often appear in tasks where sequence structure is central. It recorded an MCC score of 91.17 for detecting promoters in 300-base-pair DNA sequences. MCC measures classification quality while accounting for imbalanced classes.

The model also reported 96.03 AUC on an RNA modification task and 91.46 percent accuracy for noncoding RNA function. For protein stability, it reached a Spearman correlation of 70.63. These metrics measure different capabilities and should not be combined into one general intelligence score.

The results were not uniformly dominant. Intern-S1-Pro led on tasks including protein fluorescence, enzyme classification, and one RNA isoform benchmark. Biology-Instructions also edged ShenZhen on enhancer activity and one cross-modal sequence task.

That uneven pattern makes the release more credible than a blanket claim of superiority. It shows that architectural specialization can produce large gains on some tasks without eliminating the advantages of scale elsewhere.

The pressure therefore falls on teams building very large scientific foundation models. They must show that additional parameters deliver capabilities unavailable from smaller systems with better scientific representations. Benchmark breadth, transfer performance, and research utility will matter more than model size alone.

The same pressure applies to narrowly specialized projects. If a unified model approaches their accuracy while simplifying deployment, specialists must justify their operational complexity. Their strongest defense will be experimental validation, domain depth, and reliability under real laboratory conditions.

Native Scientific Encoders Are the Real Mechanism

ShenZhen’s central idea is architectural division of labor, not universal tokenization.

General-purpose multimodal models usually project text, images, and sometimes audio into a shared representation space. That strategy works because those modalities contain patterns that large neural networks can align at scale. Scientific data creates a harder problem.

A protein sequence is not simply a sentence with a different alphabet. Its amino acids influence folding, stability, binding, and function through relationships that span multiple scales. A molecule is also more than its written SMILES representation, because its graph structure determines chemical connectivity.

Weather presents another mismatch. A global atmospheric state contains physical variables distributed across latitude, longitude, altitude, and time. Flattening that field into a generic sequence risks discarding spatial relationships that forecasting systems need.

ShenZhen addresses the mismatch by keeping specialized front ends. Each encoder first captures the structure native to its domain. The shared Qwen backbone then operates over representations aligned with natural-language instructions and other modalities.

The decoder performs the reverse process. For weather, it generates a sequence of global atmospheric states. For medical imaging, it converts a text prompt and image into a segmentation mask. For molecules, it can return a valid SMILES representation.

This arrangement resembles a team of specialists working through one coordinator. The coordinator provides a common interface and shared reasoning space. The specialists protect details that would be damaged by forcing every problem into the same format.

The reported molecule tests illustrate the potential benefit. On six SMolInstruct property tasks, ShenZhen reportedly led or tied for first on four. It achieved 96.95 percent accuracy on blood-brain barrier penetration and 97 percent on an HIV-related classification task.

For ESOL, a benchmark that estimates water solubility, ShenZhen posted a root mean squared error of 0.550. Lower values indicate predictions closer to the reference data. The cited LlaSMol baseline recorded 1.150.

These are benchmark results, not evidence that generated molecules will become safe drugs. Property prediction usually represents one early filter in a much longer discovery process. Laboratory synthesis, toxicity testing, pharmacokinetics, and clinical evaluation remain necessary.

The weather component follows the same native-output principle. Given one initial atmospheric field, ShenZhen generates another state every six hours for up to ten days. The academy evaluated these predictions on global ERA5 data at 0.25-degree resolution.

At day ten, the release reports lower errors than ECMWF HRES for 500-hectopascal geopotential height, two-meter temperature, and mean sea-level pressure. HRES is the European Centre for Medium-Range Weather Forecasts’ operational high-resolution numerical weather prediction system.

The model card gives approximate day-ten errors of 740 versus 810 for geopotential height. It lists temperature errors near 2.65 Kelvin versus 2.90, and pressure errors around 680 Pascals versus 745.

Those figures should not be read as proof that ShenZhen outperforms an operational forecasting service. The comparison uses an offline ERA5 evaluation and a limited set of variables. Operational forecasting includes data assimilation, ensemble uncertainty, frequent updates, and many regional performance requirements.

The point is narrower but still notable. One shared model can produce a plausible ten-day atmospheric forecast after also learning biological and chemical tasks. Specialized routing appears to preserve enough weather structure for competitive offline results.

That mechanism is the article’s main reversal. Scientific AI does not necessarily become more general by stripping away domain structure. ShenZhen attempts to gain generality by preserving that structure before aligning it through a shared backbone.

Medical Imaging Shows Both the Promise and the Verification Gap

The medical imaging result is ShenZhen’s clearest practical claim and its clearest reason for caution.

The academy tested text-prompted segmentation across 102,855 image-prompt pairs from the BiomedParse test splits. The collection covers nine imaging modalities, including CT, MRI, pathology, X-ray, endoscopy, and retinal imaging.

ShenZhen achieved a mean Dice score of 91.20, according to the model card. Dice measures the overlap between a predicted region and a reference annotation. A higher percentage indicates closer agreement.

The reported comparison included seven methods. BiomedParse scored 90.73, MedSAM reached 83.55, and the original Segment Anything Model recorded 71.29. ShenZhen ranked first across the combined test set.

Its performance varied by modality. It reportedly reached 98.02 on X-rays, 93.36 on CT images, and 85.29 on MRI scans. BiomedParse remained slightly ahead on MRI, optical coherence tomography, X-ray, and fundus images.

On pathology images, ShenZhen reported 87.29, compared with 81.57 for BiomedParse. That gap is interesting because pathology images often contain complex textures and tissue boundaries. However, one benchmark cannot establish clinical reliability.

The model uses a SAM-based path for segmentation. SAM refers to Meta’s Segment Anything architecture, which converts visual prompts into object masks. ShenZhen adds scientific routing and natural-language interaction around that image component.

A researcher might request the region corresponding to a tumor, organ, or tissue structure. The system then returns a mask rather than only describing the image. That output can support quantitative analysis, annotation assistance, or preprocessing for another model.

Yet medical segmentation scores carry several limitations. Benchmark images can differ from data collected at another hospital, scanner, or patient population. Performance can degrade when imaging protocols, disease prevalence, annotation practices, or prompt wording change.

Dice also measures overlap, not clinical consequences. A high average score can hide small but important errors near a lesion boundary. It can also conceal weak performance on rare conditions or underrepresented patient groups.

The release does not provide prospective clinical validation. It also does not establish whether clinicians can use the system safely without extensive supervision. Readers should treat ShenZhen as a research model, not a diagnostic product.

Licensing adds another constraint. The package contains third-party segmentation weights covered by Meta’s SAM license and acceptable-use restrictions. The presence of an Apache license elsewhere does not automatically remove those obligations.

Independent reproduction is therefore the immediate test. Outside teams need to run the published code against untouched datasets and report their environments. They should also examine calibration, failure cases, prompt sensitivity, and cross-institution performance.

The same standard applies beyond medical images. Biological scores need validation on datasets protected against training overlap. Molecule outputs need chemical validity checks, while weather forecasts need operational comparisons across seasons and regions.

Open weights make these tests possible, but openness does not guarantee reproducibility. Researchers still need exact preprocessing steps, data splits, evaluation scripts, and suitable hardware. Missing details can materially change a benchmark result.

Scientific teams adopting the model will also face knowledge-management challenges. Reproducing a result requires tracking prompts, datasets, model versions, licenses, and failed experiments. A searchable technical knowledge base can help preserve that evidence across a research group.

This verification gap is not a reason to dismiss the release. It is the main reason an open release matters. ShenZhen’s value will become clearer when independent users discover where its unified architecture works and where specialists remain necessary.

Specialized Models Still Set the Scientific Standard

ShenZhen competes with a portfolio of specialized systems, not with one universal rival.

In genomics, Google DeepMind’s AlphaGenome focuses on predicting how DNA variants affect gene regulation. It processes sequences up to one million base pairs and produces thousands of molecular predictions. That depth differs from ShenZhen’s broader cross-domain objective.

Google first offered AlphaGenome through an API for noncommercial research. A subsequent Nature paper described its performance on regulatory variant prediction and released research resources. The AlphaGenome research shows how much domain-specific engineering can fit inside one scientific area.

In protein science, ESM3 models sequence, structure, and function together. Its developers trained the largest version with 98 billion parameters and described it as a generative model for protein design. The system remains specialized around the biology of proteins.

The ESM3 study reported the generation of a fluorescent protein distant from known examples. That type of experimental case matters because it moves beyond benchmark prediction toward a physical result.

Weather forecasting has its own specialists, including systems from Google DeepMind, Huawei, Nvidia, and ECMWF. These models increasingly compete with numerical weather prediction on selected variables and lead times. Operational agencies still combine multiple systems and human expertise.

Medical imaging has similarly diverse foundation models. MedSAM focuses on segmentation, while BiomedParse connects images and text prompts across biomedical modalities. Clinical organizations may prefer these focused systems when documentation and validation match a specific workflow.

ShenZhen’s advantage is not obvious superiority over these specialists. Its appeal lies in crossing the boundaries between them. A research project might connect genomic sequences, protein function, candidate molecules, medical images, and environmental conditions.

Today, such projects often require several models with incompatible input formats. Researchers must maintain separate environments and translate outputs between tools. Each handoff introduces technical work and potential information loss.

A shared natural-language interface can reduce those barriers. More importantly, a shared representation space might allow information learned in one modality to improve another. That cross-domain transfer remains a hypothesis until experiments isolate and measure it.

The academy’s surrounding platform strengthens this systems argument. SAIS introduced Dasheng as an agent that can break down research tasks and coordinate models and tools. Its NovaInspire platform gathers scientific models, data, papers, and patent records.

According to an official Shanghai overview, SAIS was established in 2023 by Shanghai’s municipal government and Fudan University. The institute has also developed models for materials, drug discovery, weather, and climate.

ShenZhen consolidates parts of that earlier portfolio without making every specialized model obsolete. The model’s own design still contains specialized encoders and decoders. Unification happens at the orchestration and representation layers, not through complete architectural uniformity.

That distinction should guide comparisons. The choice is not simply one general model versus many experts. The more relevant contest is between a coordinated shared backbone and a loosely connected collection of independent systems.

Large models such as Intern-S1-Pro offer another route. They seek breadth through parameter scale, extensive scientific pretraining, and multimodal reasoning. ShenZhen seeks breadth through domain-aware routing around a smaller backbone.

Both routes still need specialist tools. A trillion-parameter reasoning model may call external simulators or databases. ShenZhen relies on dedicated components inside its package. The practical question is where specialization should sit within the stack.

If ShenZhen transfers effectively across domains, it will strengthen the shared-backbone route. If independent tests reveal negative transfer, where learning one domain harms another, modular specialist systems will retain an important advantage.

Three Signals Will Decide Whether ShenZhen Changes Scientific AI

The next phase is about reproducibility, cross-domain transfer, and adoption rather than another headline benchmark.

The first signal is independent benchmark reproduction. Research groups should rerun the 20 biological tasks, six molecule tests, weather evaluation, and medical segmentation benchmark. Consistent results would strengthen the academy’s scale-efficiency argument.

Reproduction should include tests for training-data overlap. Scientific benchmarks often rely on public datasets that can enter pretraining corpora. Strong scores mean less when a model has already encountered identical sequences, labels, images, or closely related examples.

A convincing independent study would use held-out temporal data or newly collected samples. For biology, that might include recently deposited sequences. For weather, researchers can test forecasts from periods unavailable during training.

The second signal is evidence of genuine cross-domain transfer. ShenZhen’s architecture assumes that a shared backbone captures relationships useful across scientific modalities. The current benchmark set mainly measures performance within individual domains.

Researchers should compare the unified model with versions trained separately on each modality. If the shared model performs better or learns from fewer examples, that would support the academy’s central thesis. Similar or weaker performance would suggest that integration mainly simplifies deployment.

Cross-modal studies would be even more informative. One test could examine whether protein representations improve RNA-protein interaction prediction. Another could connect molecular generation with biological target information instead of evaluating isolated property benchmarks.

The third signal is sustained adoption through the open repositories and NovaInspire. Downloads alone offer a weak measure, but outside contributions reveal whether the package is usable. New evaluation scripts, domain adapters, bug reports, and reproducible experiments would indicate a functioning research community.

Adoption also depends on compute requirements. An 11-billion-parameter model is smaller than Intern-S1-Pro, but it is not lightweight for every laboratory. The package combines several large scientific components, which can complicate memory use and deployment.

Clear hardware profiles, quantized versions, and modular loading would expand access. Researchers should be able to load only the components required for one workflow. Otherwise, a unified package can become less efficient than the separate tools it seeks to replace.

Watch how specialists respond as well. If teams behind genomic, protein, weather, or imaging models publish stronger cross-domain interfaces, ShenZhen’s orchestration advantage will narrow. If they remain isolated, a shared scientific backbone becomes more attractive.

The academy should also publish fuller training details and controlled ablations. An ablation removes one component to measure its contribution. These studies can show whether gains come from the encoders, joint training, the Qwen backbone, or benchmark-specific adaptation.

The Shanghai Academy of AI for Science ShenZhen multimodal foundation model has already made one useful contribution. It turns a debate about scientific model scale into a testable architectural question.

Can a smaller backbone coordinate native scientific representations without sacrificing specialist accuracy? The open release lets research teams investigate that question directly.

The best next step is not to accept every reported score or dismiss the model because verification is incomplete. Researchers should reproduce one relevant task, document each dependency, and publish the failures alongside the successes. That evidence will determine whether ShenZhen becomes shared infrastructure or remains an ambitious collection of benchmark results.