Intern-S2-Preview – Shanghai AI Lab Open Scientific Multimodal LLM

⚠️ Disclaimer: The model reviewed in this article is a preview version that has not been officially released. All evaluations are based on the current public preview version and do not represent the final release. This content is for technical reference only and does not constitute any recommendation or guarantee.

1. What Is Intern-S2-Preview

Intern-S2-Preview is Shanghai AI Laboratory's open scientific multimodal LLM preview, delivering trillion-class scientific capability at 35B parameters. Its "general–specialist fusion" training pipeline plus RL delivers the first open general LLM with material crystal structure generation—MolecularIQ 57.26, crystal pass rate >40%, ahead of GPT-5.5-class closed models on key tasks. It supports complex scientific reasoning, multi-omics biology, and agent execution, with Ascend Atlas 900 A3 co-design for algorithm–system–compute efficiency.

Technical positioning and domain: A scientific multimodal foundation model fusing hundreds of specialist tasks into one general architecture—molecular reasoning, crystal generation, omics, remote sensing—unlike narrow science models or generic chat LLMs alone.

Development background: Building on InternLM, the lab saw science workloads either needing trillion-scale params or single-domain tools. Intern-S2-Preview targets breakthrough scientific ability at deployable 35B scale under the "Shusheng" (Intern) roadmap from general to specialist AI for Science.

Core value: It lowers the dual barrier of deployment cost and weak open scientific skills. 35B fits modest GPU clusters while scoring strongly on MolecularIQ, crystal generation, SciCode, and agent benches—plus coordinate regression for crystals without diffusion-only pipelines.

Technical characteristics: End-to-end general–specialist fusion from pretrain through RL with task scaling; Fourier positional encoding (FoPE) and real-valued prediction heads for structural coordinates; XTuner + LMDeploy train/serve stack optimized for Ascend Atlas 900 A3.

2. Key Features

• Crystal structure generation: First open general LLM with material crystal generation via real-valued coordinate prediction—no diffusion dependency; MolecularIQ 57.26, >40% pass rate.

• Scientific multimodal understanding: Omics sequences, microscopy QA, molecular instructions, structure reasoning, remote sensing, and composite scientific QA across modalities.

• Math and complex reasoning: IMO-level math, multimodal math, competition tasks with compressed chain-of-thought; IMO-Bench 84.00, MMLU-Pro 88.00.

• Long-context text and multimodal reasoning: Handles long papers and cross-modal scientific inputs for literature synthesis and experiment planning.

• Scientific chart and data QA: Interprets plots and tables in papers with domain-aware answers.

• Scientific code generation: SciCode 39.64—Python/Fortran/MATLAB-style research scripts and numerics.

• General agent execution: Scientific agents, OpenClaw coding agents, software engineering tasks—SGI-Bench 52.52, PinchBench 88.22.

• Instruction following and general coding: High-precision tool use and code production for automated lab workflows.

3. How to Use

1. Online chat: https://chat.intern-ai.org.cn — test science QA, structure tasks, and codegen without setup.

2. API: ChatAPI with model: intern-s2-preview; keep thinking_mode on for reasoning and agents; streaming supported.

3. Agents: For OpenClaw-style harnesses, leave thinking mode enabled for stable decomposition and tool calls; review official agent examples first.

4. Local deploy: Weights on HuggingFace/ModelScope; LMDeploy recommended (~140 GB BF16, ~4× A100 80GB or Ascend Atlas 900 A3). Example: lmdeploy serve internlm/Intern-S2-Preview.

5. Fine-tuning: XTuner with RL/MTP shared weights; ~8× A100 suggested, lr ~1e-5, batch size by VRAM.

6. Crystal tasks: Prompt with composition or structure description; validate outputs with VASP/Quantum ESPRESSO DFT.

4. Pros and Cons

5. Comparison with Similar Tools

Selection guidance: Material science, molecular bio, and drug discovery teams should default to Intern-S2-Preview for crystals and agents. Pure SciCode-heavy coding may slightly favor Step3.5-Flash but without crystal skills. Budget-conscious teams gain the best open science bundle at 35B.

6. Editor's Take

Intern-S2-Preview is a milestone in AI for Science: specialist tasks baked into pretrain–RL scaling, not bolted on later. Open crystal coordinate regression in a general LLM is practically meaningful for materials workflows—not just a leaderboard trick.

35B deployability on modest clusters vs. trillion gossip is the other headline. Labs can run structure prediction, omics QA, and agents without hyperscaler budgets—if they can access A100-class hardware or APIs.

As a preview, docs and community plugins will grow; crystal outputs still need DFT validation. Even so, fusion training and Ascend optimization make this a reference point for open scientific foundations.

— Innovation 5/5, utility 4.5/5, ease 4/5, ecosystem 3.5/5.

7. Use Cases

• Molecular biology and drug discovery: Omics analysis, protein/molecule reasoning, target and compound triage.

• Materials discovery: Candidate crystal generation from compositions—accelerate semiconductor, catalyst, and battery research vs. pure trial-and-error.

• Scientific computing: Auto-generate simulation prep scripts (MD, DFT pipelines, bioinformatics).

• Microscopy and remote sensing: Image QA from lab slides to satellite imagery.

• Advanced math and teaching: Competition-level proofs and step-by-step reasoning.

• Research agents: Literature search, experiment design, analysis code, and report drafts in multi-step workflows.

8. FAQ

Q: How does Intern-S2-Preview differ from InternLM?

A: InternLM is general-purpose; Intern-S2-Preview extends the Shusheng line with fused scientific training—crystals, molecules, omics, and agents at 35B.

Q: Chinese support?

A: Scientific Chinese works well; casual chat/creative Chinese may trail dedicated Chinese chat models. Mixed CN/EN technical prompts are fine.

Q: Crystal accuracy—trust without DFT?

A: Use model outputs as hypotheses; validate with VASP/QE DFF. Pass rates are strong vs. closed baselines but not a substitute for physics verification.

Q: VRAM requirements?

A: ~140 GB BF16 full weights on ~4× A100 80GB; INT4 ~70 GB with some quality loss. Consumer GPUs need API access.

Q: Commercial use?

A: Open license on HuggingFace—check LICENSE for details and legal review for regulated industries.

Q: Domain fine-tuning?

A: XTuner on domain papers/data, lr ~1e-5, 8× A100, 10–20 epochs typical starting point.

Q: Multi-turn memory?

A: Supported within context limits; batch very long literature to avoid truncation.

9. Project Links

• Online demo: https://chat.intern-ai.org.cn/
• HuggingFace: https://huggingface.co/internlm/Intern-S2-Preview
• ModelScope: https://modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S2-Preview