Intern-S2-397B: A 397B Scientific-Reasoning Backbone on a Qwen3.5 MoE Base

What happened

internlm published Intern-S2-397B to the Hugging Face model hub on 13 September 2026. It is a 397B-parameter, image-text-to-text multimodal model on a Qwen3.5 MoE backbone, Apache-2.0 licensed, and positioned by the team as a foundation model for scientific intelligence and long-horizon agents.

Context

The qwen3_5_moe hub tag tells you the backbone is not from scratch: it is a Qwen3.5 mixture-of-experts derivative, so 397B is total parameters across experts and the active count per forward pass will be lower. The model card calls it the team's "most capable multimodal foundation model for scientific intelligence and long-horizon agents." That framing is deliberate. Pretraining runs on raw pages of scientific literature without intermediate parsing, RL spans more than 20 scientific domains, and agent training uses sandboxed environments for black-box agentic RL. This is a reasoning backbone, not a creative-generation model. The GitHub link in the card points to Intern-S1, not an S2 repo; the sources do not explain the discrepancy.

How it works

The model takes image and text in and produces text out. Visual pretraining runs directly on raw pages of scientific literature with no intermediate parsing step, so the model sees layout, figures, and equations as printed rather than as a flattened text stream. Reinforcement learning then covers more than 20 scientific domains, and a separate agent-training loop connects the model to multiple agent frameworks inside large-scale sandboxed environments.

Inference targets are LMDeploy, vLLM, and SGLang. The model exposes an OpenAI-compatible tool-calling API, demonstrated with the lmdeploy API server. Maximum inference length is 256K tokens for text reasoning and 64K tokens when multimodal input is present. Weights ship in safetensors format. Evaluation ran through OpenCompass, VLMEvalKit, and AgentCompass; the results are in an image file in the model card, not extractable numbers.

Our read

The parameter count is the least interesting number here. The Qwen3.5 MoE tag tells you the architecture ceiling is inherited, not invented. What is new is the training objective: raw scientific-page pretraining, multi-domain RL, and sandboxed agent RL form a pipeline aimed at a reasoning engine that plans, calls tools, and iterates over long horizons. That is a different product than a chatbot that answers a question and stops.

Two gaps in the model card are worth flagging. The benchmark results live in a PNG file, not in text. You cannot grep them, diff them, or feed them to a spreadsheet. For a model this size, that is a small but real trust gap. And the GitHub link points to the Intern-S1 repository, not an S2 one. The sources do not say whether that is a copy-paste error or a shared codebase, and I would want to know which before pulling the repo.

The "long-horizon agent" framing is what matters for a studio. The model is trained to operate inside sandboxed environments with tool access. That makes it a planning and scripting layer: read reference frames, output a structured shot list, critique a storyboard, explain a scientific figure for a doc-to-video pipeline. It is not a diffusion model. It will not render a video. It reads an image and writes text about it, or it reasons through a 256K-token document.

What this changes

Nothing changes in the local ComfyUI node graph. The sources do not state hardware requirements for self-hosted inference, but a 397B MoE model is not a single-consumer-GPU workload; it belongs in the cloud API tier. Integration means a custom ComfyUI node that calls the model's OpenAI-compatible API, whether you self-host on vLLM or SGLang or use Hugging Face's hosted endpoints.

Where it fits: a reasoning and script layer. Feed it reference frames or a scientific figure as image input, get structured text back. The 256K-token text context handles a full script in one pass; 64K is the ceiling when images are in the prompt. Apache-2.0 clears commercial use of client-facing output without royalty. Operationally, this is a new external API dependency to budget and rate-limit, not a replacement for any node already in the pipeline.

License

Apache-2.0, as stated on the Hugging Face model card. Commercial use is permitted with no royalty and no usage-count ceiling. You can ship client-facing output generated by this model without additional licensing negotiation.

Key takeaways

• Intern-S2-397B is a 397B-parameter, image-text-to-text multimodal model on a Qwen3.5 MoE backbone, published by internlm under Apache-2.0 on 13 September 2026.
• It is a reasoning and scripting layer, not a generation model: it reads images and text and produces text output. It will not generate video or images.
• The training pipeline (raw scientific-page pretraining, 20+ domain RL, sandboxed agent RL) targets long-horizon, tool-using workloads rather than single-turn Q&A.
• Deployment is via vLLM, SGLang, or LMDeploy; the model exposes an OpenAI-compatible tool-calling API. It is not a native ComfyUI backend.
• Benchmark scores are presented as an image in the model card; no numeric results are extractable from the text. The GitHub link references the Intern-S1 repository, not S2.

Sources

1. internlm/Intern-S2-397B — image-text-to-text on Hugging Face — tier 1