Intern-S1
Open-source multimodal reasoning model for science
Intern-S1 is an open-source multimodal reasoning model released by the InternLM team. It combines a large language backbone with a vision encoder to handle scientific tasks. The system processes text, images and video using specialised tokenisers for molecular data.
- Built on 235B MoE language model and 6B vision encoder
- Trained on 5 trillion tokens with 50% scientific data
- Top scores on ChemBench, MatBench and MathVista
- Supports text, image and video inputs natively
- Dynamic tokenizer for molecular and protein sequences
What it is
Intern-S1 is a multimodal reasoning model designed for general tasks and specialised scientific work. The InternLM team built it on a 235 billion parameter Mixture of Experts language model paired with a six billion parameter vision encoder. It processes text, images and video inputs. The system excels at interpreting chemical structures, understanding protein sequences and planning synthesis routes. It serves as a research assistant for complex scientific applications.
How it came about
The model was introduced in August 2025 via a technical report and public code repositories. It represents the most advanced open-source release from the InternLM group to date. The team positioned it to rival closed-source commercial systems in scientific domains. Distribution occurs through Hugging Face collections and GitHub repositories alongside an online chat interface.
Inside the model
The architecture uses a Qwen3 based Mixture of Experts backbone with an InternViT vision component. Training involved five trillion tokens of multimodal data. Over half of this corpus consisted of specialised scientific domain tokens. A dynamic tokenizer allows native processing of molecular formulas and seismic signals. This design retains general capabilities while embedding deep expertise in chemistry and physics.
What it can do
Benchmarks show strong performance in scientific reasoning tasks. It achieved top scores among all models on ChemBench, MatBench and MathVista. The system also leads open-source models on MMMU and GPQA evaluations. General benchmarks like MMLU-Pro show competitive results against commercial rivals. It handles tool calling for external API interactions and code execution.
How to use it
Weights are available openly under permissive terms on Hugging Face. Developers can run the model locally using Ollama or deploy it via vLLM and LMDeploy. The system requires significant hardware resources. Full precision deployment needs eight A100 GPUs. An FP8 quantised version reduces this requirement to four H800 cards. It supports OpenAI compatible API servers for integration.
Questions people ask about Intern-S1
What hardware is needed to run Intern-S1?
Full precision deployment requires eight A100 GPUs. The FP8 quantised version needs four H800 cards or two H200 units. Local deployment is possible via Ollama for smaller scale testing.
What scientific tasks does Intern-S1 handle best?
It specialises in interpreting chemical structures and understanding protein sequences. The model also plans compound synthesis routes and processes seismic signals using its dynamic tokenizer.
Is Intern-S1 open source?
Yes, the weights are openly available on Hugging Face. The code resides on GitHub. It supports standard inference frameworks like vLLM and LMDeploy for local or server deployment.
Good reading and sources
There is no English Wikipedia article on Intern-S1 yet, so the facts on this page come from the sources listed above: huggingface.co/internlm/Intern-S1. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.