Overview
Intern-S2-397B is the Shanghai Artificial Intelligence Laboratory's multimodal foundation model for scientific intelligence and long-horizon agents, published under the InternLM organisation on Hugging Face on 13 September 2026 with an Apache-2.0 licence. The model card introduces it as the lab's most capable model of this kind, scaled along three axes at once: pre-training, reinforcement-learning task coverage, and interactive agent environments.
The pre-training idea is visual. Rather than parsing scientific literature into text first, the model learns directly from raw pages, jointly modelling symbolic semantics and visual relationships in a shared representation space. The lab argues this preserves text–visual correspondence, strengthens spatial and visual reasoning, and improves data efficiency — a reasonable claim to test if your corpus is full of equations, plots and figures that lose meaning once flattened.
Post-training scales reinforcement learning across more than 20 scientific domains trained jointly, including specialised tasks such as biomolecular interaction design and material structure generation, plus black-box agentic reinforcement learning in large-scale sandboxed environments connected to multiple agent frameworks. The card reports leading general-reasoning results among open-source models, with the comparison published as a figure rather than a table — so this page carries no transcribed benchmark numbers.
Mechanically it is a 397B Mixture-of-Experts with 512 experts and 10 active per token across 60 layers, built on the Qwen3.5-MoE conditional-generation architecture, with linear attention interleaved with full attention every fourth layer and a 262,144-token position limit. Thinking mode is on by default. It serves through LMDeploy, vLLM and SGLang, exposes OpenAI-compatible endpoints with tool calling, and can be pointed at from Claude Code; BF16 and FP8 checkpoints are on both Hugging Face and ModelScope, and there is a hosted chat at chat.intern-ai.org.cn.
| Released | 2026-09-13 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 397B Mixture-of-Experts — 512 experts, 10 active per token, 60 layers |
| Context | 262,144 tokens (max_position_embeddings); evaluated at 256K for text reasoning and 64K for multimodal benchmarks |
| Architecture | Multimodal Mixture-of-Experts (Hugging Face architecture Qwen3_5MoeForConditionalGeneration): hidden size 4,096, 32 attention heads with head dimension 256, linear attention with full attention every fourth layer, and one multi-token-prediction layer |
| Modalities | Text, Vision, Time series |
| Status | Open weights — BF16 and FP8 checkpoints published on Hugging Face and ModelScope |
Strengths
- Apache-2.0 on a 397B multimodal MoE, in both BF16 and FP8, on Hugging Face and ModelScope
- Pre-trained directly on raw pages of scientific literature, keeping text and figures in one representation space
- Reinforcement learning across more than 20 scientific domains, including biomolecular interaction design and material structure generation
- Long-horizon agent training in sandboxed environments wired to multiple agent frameworks
- 262,144-token position limit, evaluated at 256K for text reasoning and 64K for multimodal tasks
- Serves on LMDeploy, vLLM and SGLang with OpenAI-compatible tool calling
Best for
- Reading scientific papers as documents — equations, plots and figures together rather than stripped to text
- Domain-science tasks such as biomolecular interaction design or material structure generation
- Long-horizon research and engineering agents that need tool calls and sandboxed execution
- Self-hosted deployments where an Apache-2.0 licence on a frontier-scale multimodal model is a requirement
How to access
| Provider | Model ID |
|---|---|
| Intern chat (hosted) ↗ | — |
FAQ
What makes Intern-S2-397B's pre-training different?
It learns from raw pages of scientific literature rather than from parsed text. The model card calls this a new vision-language pre-training paradigm that jointly models symbolic semantics and visual relationships in a shared representation space without intermediate parsing, preserving text–visual correspondence.
What is the context window?
The published config sets max_position_embeddings to 262,144. The model card notes the evaluation settings: a maximum inference length of 256K tokens for text reasoning benchmarks, and 64K tokens for multimodal benchmarks.
Which serving frameworks are supported?
LMDeploy, vLLM and SGLang, per the model card's deployment section. The card also documents tool calling through an OpenAI-compatible API served by the LMDeploy API server.
Where do the published benchmark results live?
The model card publishes them as a figure (intern-s2-397b.jpg) rather than a table, evaluated with OpenCompass, VLMEvalKit and AgentCompass. Because the numbers are only legible inside an image, none are transcribed here.
How does it relate to Intern-S1 and the Intern-S2 previews?
Intern-S1 was published in July 2025 and Intern-S2-Preview-397B in July 2026, both Apache-2.0 under the same InternLM organisation; Intern-S2-397B followed in September 2026, alongside an FP8 checkpoint. The project shares the InternLM/Intern-S1 GitHub repository.