InternLM/Intern-S1

A Scientific Multimodal Foundation Model

What it solves

Intern-S series provides a collection of multimodal foundation models designed specifically for scientific intelligence and long-horizon agentic tasks. It addresses the gap in open-source models' ability to handle complex scientific reasoning, interpret specialized modalities (like molecular formulas and protein sequences), and execute multi-step scientific workflows.

How it works

The series employs several scaling and training strategies across different model versions:

  • Visual Pre-training: Models like Intern-S2-Preview-397B learn directly from raw scientific literature pages, modeling symbolic semantics and visual relationships without intermediate parsing.
  • Reinforcement Learning (RL): The models use large-scale multi-task RL across 20+ scientific domains and black-box agentic RL in sandboxed environments to improve general reasoning and long-horizon task execution.
  • Architecture Variations: The series includes trillion-scale Mixture-of-Experts (MoE) models (Intern-S1-Pro) using STE routing and grouped routing for stability, as well as efficient medium-sized models (Intern-S2-Preview-35B) utilizing shared-weight MTP and CoT compression to speed up token generation.
  • Specialized Modalities: The models use dynamic tokenizers and Fourier Position Encoding (FoPE) to natively understand molecular formulas, protein sequences, and long heterogeneous time-series signals.

Who it’s for

  • Scientific Researchers: Those needing AI assistants for chemistry, materials science, life sciences, and earth sciences.
  • AI Developers: Engineers building long-horizon agents for scientific workflows.
  • ML Researchers: Those studying multimodal scaling and reinforcement learning for specialized domains.

Highlights

  • Broad Scientific Coverage: Specialized capabilities in biomolecular interaction design, material crystal structure generation, and compound synthesis planning.
  • High Efficiency: The 35B model achieves performance comparable to trillion-scale models on core scientific tasks through task scaling.
  • Massive Training Data: Intern-S1 was pretrained on 5 trillion tokens, with over 2.5 trillion dedicated to scientific domains.
  • Diverse Model Zoo: Offers a range of sizes from lightweight (Intern-S1-mini) to trillion-parameter MoE (Intern-S1-Pro).

Related

  • Project
    OpenGVLab/InternVideo

    A series of video foundation models and large-scale datasets designed for multimodal video understanding, long-context modeling, and contextual reasoning.

  • Project
    MMMU-Benchmark/MMMU

    A massive multi-discipline benchmark for evaluating multimodal AI models on college-level subject knowledge and complex reasoning across 30 different subjects.

  • Project
    SkyworkAI/Skywork-R1V

    Skywork-R1V is an open-source multimodal reasoning model that uses reinforcement learning and visual chain-of-thought to achieve state-of-the-art performance in complex visual logic, math, and physics tasks.

  • Project
    microsoft/unilm

    A comprehensive collection of foundation models and architectures from Microsoft focusing on large-scale self-supervised pre-training across text, vision, speech, and multimodal tasks.

  • Project
    zai-org/GLM-4

    A series of open-source LLMs featuring 9B and 32B parameter models with specialized variants for general chat, deep reasoning, and complex rumination tasks.