Refreshing the Frontiers of Technology! InternLM Launches a Lightweight Multimodal Reasoning Model with 8B Parameters

[baoshi.rao]

Recently, the InternLM team officially released its open-source lightweight multimodal reasoning model—Intern-S1-mini. With only 8B parameters, this model combines the advanced Qwen3-8B language model with the 0.3B visual encoder InternViT, demonstrating powerful processing capabilities and flexibility.

Intern-S1-mini has undergone large-scale pre-training, utilizing over 5 trillion tokens of data. Notably, more than 2.5 trillion tokens come from various scientific fields such as chemistry, physics, biology, and materials. This enables Intern-S1-mini to not only handle conventional text and visual inputs but also parse complex molecular formulas, protein sequences, and effectively plan synthesis pathways, showcasing its broad application potential in scientific research.

According to the benchmark test results provided by the official, Intern-S1-mini outperforms similar models in multiple domains. On tasks such as MMLU-Pro, MMMU, GPQA, and AIME2024/2025, the model's performance is astonishing, with a ChemBench score of 76.47, MatBench score of 61.55, and ProteinLMBench score of 58.47. These results not only prove the model's robust capabilities but also highlight its compatibility with text, image, and video inputs.

Interestingly, Intern-S1-mini defaults to enabling "thinking mode," which users can toggle with a simple command (enable_thinking). This design enhances the model's interactivity, offering users a more flexible experience.

In today's rapidly advancing technological landscape, the release of Intern-S1-mini undoubtedly provides researchers and developers with a new tool to drive innovation and breakthroughs in the field of multimodal reasoning. Whether in fundamental research or practical applications, this model will undoubtedly be a focal point worth watching.