1、Page 1Shanghai Artificial Intelligence LaboratoryPage 2Shanghai Artificial Intelligence LaboratoryPage 3Shanghai Artificial Intelligence LaboratoryPage 4Shanghai Artificial Intelligence LaboratoryPage 5Shanghai Artificial Intelligence LaboratoryPage 6Shanghai Artificial Intelligence Laboratory高价值区域:
2、AGI4S通专融合技术架构(SAGE)科学发现的制高点泛化性专业能力路线AlphaFold3DeepBlueAlphaGoDALL-E 22023年提出AlphaFoldLife ScienceAlphaProofEvo 2MatterGen局限性:难以超越人类顶尖 10%专家水平MathematicsLife ScienceMaterials三大缩放定律(Three Scaling Laws)From:Jensen Huang at NVIDIA GTC2025Pre-Training Scaling数据受限,增速放缓Post-Training Scaling基于监督微调(SFT)/人类反馈
3、强化学习(RLHF)Test-Time Scaling长思考/推理GPT-4DeepSeek R1Anthropic Claude 4非专业人才水平专业人才平均水平专业人才Top 10%水平专业人才Top 1%水平全面超越人专业性通专融合能力发展路径Page 7Shanghai Artificial Intelligence Laboratory4T tokenInternLM318T tokenMainstream LLMInternLM(IQPT,Intelligence Quality per Token)具身交互、规划InternVLInternThinker高 IQPT 与混合推理能
4、力高性能开源视觉语言大模型(VLM)具备卓越推理能力的通专融合模型大规模可验证交互环境交互反馈高密度监督交互学习多模态感知、理解和推理自定义开放任务专业模型通用模型世界模型Intern-S1/Intern-S1-Pro全球领先的科学多模态大模型Intern 系列模型:语言、视觉、科学领域全覆盖Page 8Shanghai Artificial Intelligence Laboratory一、概述Page 9Shanghai Artificial Intelligence Laboratoryn模型架构多模态混合专家模型(MoE),128 个专家,总参数量 241Bn通用能力在文本与多模态通用
5、能力评测基准上,性能超越 InternVL3 和 Qwen2.5-VLn科学能力在化学、生命科学、地球科学等 9 项科学评测基准中,8 项性能超越 Grok-4Intern-S1Intern-S1Time-SeriesMolecular StructureFeatures for Materials DesignGravitational waves analysisPage 10Shanghai Artificial Intelligence LaboratoryIntern-S1Open-source modelsProprietary models全能高手+科学明星Intern-S1In
6、ternVL3-78BQwen3-VLGrok-4Gemini-2.5-ProGPT-5GLM-4.5VClaude Sonnet435404550556065505560657075808590050100Intern-S1InternVL3-78BQwen2.5-VL-72BDS-R1-0528Qwen3-235B-A22BKimi-K2-InstructGemini-2.5-ProGrok-4GPT-5MMLU-Pro020406080100Intern-S1InternVL3-78BQwen2.5-VL-72BDS-R1-0528Qwen3-235B-A22BKimi-K2-Instr