InternLM open-sources Intern-S2-397B multimodal model for scientific and agentic tasks
InternLM released Intern-S2-397B, a multimodal foundation model for scientific intelligence and long-horizon agents, on Hugging Face under Apache-2.0. The model supports image-text-to-text tasks and can be deployed with LMDeploy, vLLM, or SGLang.
China context
- Original name
- 上海人工智能实验室
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can integrate Intern-S2-397B into agent frameworks via OpenAI-compatible or Anthropic-compatible endpoints, or self-host with LMDeploy, vLLM, or SGLang.
- For investors
- The release of a high-capability open-weights model by a leading Chinese lab may pressure proprietary model providers and accelerate adoption of open-source alternatives in scientific and enterprise AI.
InternLM released Intern-S2-397B, a multimodal foundation model for scientific intelligence and long-horizon agents, on Hugging Face under Apache-2.0. The model supports image-text-to-text tasks and can be deployed with LMDeploy, vLLM, or SGLang. It features a new vision-language pre-training paradigm, scientific modality reasoning and generation across more than 20 domains, and general and scientific long-horizon agent capabilities.
Intern-S2-397B scales pre-training, reinforcement-learning task coverage, and interactive agent environments. It uses a vision-language pre-training paradigm that learns directly from raw pages of scientific literature, jointly modeling symbolic semantics and visual relationships. The model supports tool calling, thinking/non-thinking modes, time series inference (LMDeploy only), and integration with agent frameworks via OpenAI-compatible or Anthropic-compatible endpoints.
Developers outside China can now access and deploy a state-of-the-art open-weights multimodal model for scientific and agentic tasks, reducing reliance on proprietary APIs and enabling customization for specialized domains.
Organizations can self-host Intern-S2-397B to build scientific AI applications and long-horizon agents without per-token API costs, potentially lowering operational expenses for high-volume inference.
Next signals to watch include independent benchmark evaluations of Intern-S2-397B against other open-source models, adoption in scientific research workflows, and the release of fine-tuned variants or smaller versions.