Summary Box [In a hurry? Just read thisā”]
- Shanghai AI Laboratory released Intern-S1-Pro, a massive open-source 1 trillion parameter MoE multimodal model with 22B active parameters.
- It excels in scientific reasoning and AI4Science tasks, often outperforming or matching top closed-source models on specialized benchmarks.
- Key innovations include STE Routing, Grouped Routing, FoPE (Fourier Position Encoding), and advanced time-series modeling for heterogeneous signals.
- The model supports multimodal inputs (images, time-series, text) and integrates with tools like vLLM and SGLang for easy deployment.
- It shows strong performance across scientific domains (biology, chemistry, physics, earth science) and remains competitive in general knowledge, math, and agent tasks.
Shanghai AI Laboratory has launched Intern-S1-Pro, an advanced open-source multimodal large language model with a massive 1 trillion parameters in a Mixture-of-Experts (MoE) architecture.
This release, announced in early February 2026, positions the model as a leader in scientific reasoning and AI4Science tasks, where it competes effectively against prominent closed-source alternatives.
With an active parameter count of 22 billion, Intern-S1-Pro emphasizes efficiency, stability, and broad applicability across scientific and general domains.
Designed for high-performance multimodal tasks, the model integrates sophisticated techniques to handle complex data types, including images, time-series signals, and text.
It supports ecosystem tools like vLLM and SGLang, facilitating easy deployment and inference.
Developers can access the model weights and code through major platforms, enabling rapid integration into research and production environments.
Key Innovations Driving Performance
Intern-S1-Pro incorporates several cutting-edge features to enhance its capabilities:
- STE Routing: Enables dense gradient updates for router training, improving overall model efficiency.
- Grouped Routing: Ensures stable convergence and balanced load distribution across experts during parallel processing.
- Fourier Position Encoding (FoPE): Combined with enhanced time-series modeling, this allows precise representation of physical signals, supporting heterogeneous sequences from 10^0 to 10^6 points.
- Multimodal Integration: Strong handling of visual, textual, and sequential data for comprehensive scientific analysis.
These advancements contribute to the model’s top-tier results on advanced reasoning benchmarks, making it a valuable tool for researchers in fields like biology, chemistry, physics, and earth sciences.
Benchmark Performance on Scientific Tasks
Intern-S1-Pro demonstrates superior performance across a range of scientific benchmarks, often outperforming or matching leading models. The following table summarizes scores on key scientific tasks:
| Benchmark | Description | Intern-S1-Pro (1T-A22B) | Qwen3-VL-235B-Thinking | Kimi-K2.5 (1T-A32B) | GPT-5.2 | Gemini-3-Pro |
|---|---|---|---|---|---|---|
| SciReasoner | Scientific Reasoning | 55.5 | 11.9 | 15.3 | 13.6 | 14.7 |
| SFE | Scientific Multimodal Tasks | 52.7 | 41.4 | 53.7 | 47.5 | 58.9 |
| SmolInstruct | Small Molecule | 74.8 | 36.6 | 53.5 | 48.2 | 58.3 |
| MatBench | Materials Property Prediction | 72.8 | 49.7 | 60.0 | 53.6 | 64.9 |
| Mol-Instructions | Bio-molecular Instruction | 48.8 | 8.9 | 20.0 | 12.3 | 34.6 |
| MicroVQA | Biological Microscopy | 63.3 | 53.8 | 55.4 | 60.4 | 69.0 |
| Biology-Instruction | Multi-Omics Sequence | 52.5 | 6.2 | 10.7 | 10.2 | 12.0 |
| ZRemoteBench | Remote Sensing | 67.8 | 51.2 | 46.4 | 50.4 | 51.8 |
| MSEarth-MCQ | Earth Science | 56.2 | 52.7 | 61.9 | 62.6 | 65.8 |
Benchmark Performance on General Tasks
In general tasks, Intern-S1-Pro also shows competitive edges, particularly in knowledge, reasoning, and visual grounding:
| Benchmark | Description | Intern-S1-Pro (1T-A22B) | Qwen3-VL-235B-Thinking | Kimi-K2.5 (1T-A32B) | GPT-5.2 | Gemini-3-Pro |
|---|---|---|---|---|---|---|
| MMMU-Pro | Knowledge & Reasoning | 72.8 | 69.9 | 78.5 | 79.3 | 81.0 |
| MMLU-Pro | Knowledge & Reasoning | 86.6 | 83.4 | 87.1 | 85.9 | 89.3 |
| AIME-2025 | Math Reasoning | 93.1 | 90.0 | 96.1 | 100.0 | 95.0 |
| IMO-Answer-Bench | Math Reasoning | 77.3 | 72.3 | 81.8 | 86.3 | 81.3 |
| RefCOCO-avg | Visual Grounding | 91.9 | 91.1 | 87.8 | 54.9 | 76.2 |
| IFBench | Instruction Following | 71.2 | 58.7 | 69.7 | 75.4 | 70.4 |
| OCRBench V2 (ENG / CHN) | OCR Generation | 60.1 / 60.6 | 66.8 / 63.8 | 64.2 / 57.4 | 56.4 / 54.6 | 68.0 / 52.5 |
| SArena (Icon) | SVG Generation | 83.5 | 76.3 | 77.3 | 55.5 | 82.6 |
| Koding | Code | 74.3 | 72.0 | 85.0 | 87.7 | 86.9 |
| Coder (Text-Only) | Agent | 73.4 | 47.8 | 79.9 | 71.1 | 75.5 |
| TauĀ²-Bench | Agent | 80.9 | 57.4 | 76.8 | 76.6 | 85.4 |
| ScreenSpot V2 | Agent & Grounding | 93.6 | 92.8 | 92.4 | 49.4 | 94.7 |
Implications for AI Research and Applications
The benchmarks reveal Intern-S1-Pro‘s strengths in specialized scientific domains, where it often leads in efficiency and accuracy.
For general tasks, it holds its own against giants like GPT-5.2 and Gemini-3-Pro, particularly in math and visual tasks.
This open-source release democratizes access to powerful multimodal AI, fostering innovation in scientific discovery, materials design, and beyond.
Researchers can leverage its MoE efficiency for scalable deployments, while the advanced encodings open new possibilities for time-series analysis in physics and biology.
