SSyncDev

Intern-S2-397B

SyncDev model directory · Profile updated Nov 18, 2026

Compare with another model

Intern-S2-397B is Shanghai AI Lab's open multimodal flagship for scientific reasoning and long-horizon agents, with thinking mode on by default.

Key specifications

Specification score
54.8
Context window
262.14K
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: No API pricing stated; open weights for self-hosting

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input

Produces

  • Text output

Capabilities

ReasoningFunction callingVisionOpen weights

What is Intern-S2-397B?

Shanghai AI Laboratory calls Intern-S2-397B its most capable multimodal foundation model for science and long-horizon agents. The ambition shows in the training, which starts somewhere unusual: raw pages of scientific papers, fed in as images.

Most science models learn from parsed text, so equations, figure captions and table layouts lose their spatial relationships. Intern-S2 learns from the page itself, keeping symbols and visual layout in one representation. The September 2026 release builds on a preview that went up in July.

Three scaling bets

  1. Visual pre-training on literature, with no intermediate parsing step.
  2. Joint scientific RL across more than 20 domains, including biomolecular interaction design and material structure generation.
  3. Black-box agentic RL that connects several agent frameworks to large sandboxed environments for long tasks.

Specification

Field Value
Hub repo internlm/Intern-S2-397B (FP8 build also published)
Released 13 September 2026 (preview on 16 July)
Architecture Qwen3.5-style MoE: 512 experts, 10 active, 60 layers
Positions 262,144
Input / output Text and images / text
Serving LMDeploy, vLLM, SGLang
Licence Apache-2.0

Benchmarks were run at up to 256K tokens of inference length for text and 64K for multimodal tasks. Shanghai AI Lab compares it with recent open and closed models; those are publisher scores.

Thinking, tools and sampling

Thinking mode is on by default. You can switch it off with enable_thinking=False in the chat template, or per request through chat_template_kwargs, but the lab advises against doing so for agent work. Tool calling works through any OpenAI-compatible server; LMDeploy is the reference. Recommended sampling is temperature 0.8, top-p 0.95, top-k 50, min-p 0.

How it sits next to its siblings

The lab's own Intern-S2-Mobius is a 35B experiment in separating knowledge from reasoning, built for efficiency rather than peak scores. For large open alternatives, compare with Qwen3.5 397B-A17B and other models from Shanghai AI Lab's Hugging Face organisation.

Worth testing if

You work with papers, patents or lab data where figures matter. Feed it twenty page images with questions that require reading a plot or a chemical structure, and compare answers with a text-only pipeline on the same PDFs. Then run one long agent task end to end and watch where it gives up. Weight memory is the constraint: plan multi-GPU serving even with the FP8 build.

How to evaluate Intern-S2-397B

Compare the model on the work you need to complete.

Tasks and inputs

Intern-S2-397B is listed for Image-text-to-text, Multimodal, Text generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 262.14K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

About 397B parameters in a mixture-of-experts layout with 512 experts and 10 active per token; an FP8 build is available.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

sensetime

SenseNova-Loop-9B

SenseNova-Loop-9B is SenseTime's 9B looped multimodal model: 32 layers run as 56, 512K context, ~15 GiB at 256K in 4-bit, open weights under Apache 2.0.

524.29K context

Tencent

Hy4 Flash

Tencent Hy4 Flash is the low-cost Hy4 tier with open weights: a 220B MoE (~11.3B active) for agents and documents, with 1M context and image input.

1.05M context

shanghai-ai-lab

Sage S2.1 Flash

Sage S2.1 Flash is Shanghai AI Lab's open scientific multimodal agent model, reading molecules, crystals and proteins alongside text and images.

— context

Meituan

LongCat-2.5

LongCat-2.5 is Meituan's open-weight MIT-licensed multimodal MoE for long-horizon agents: screenshots, charts, video and multi-step coding.

— context

sensetime

SenseNova-Loop-9B-Preview

SenseNova-Loop-9B-Preview is SenseTime's looped 9B text-and-image model for long-context agentic work, running shared layers twice for extra depth.

262.14K context

Liquid AI

LFM3-VL-2B

LFM3-VL-2B is Liquid AI's open 2B vision-language model for images and video on-device, strong on documents, charts and GUI screenshots.

— context