[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-xpeng-turingvit":3,"topics-all":36,"news-related-224b056b-67e8-4488-9f5f-e56497de050d":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"224b056b-67e8-4488-9f5f-e56497de050d","小鹏 TuringViT 把视觉 Transformer 训练成本砍到一成：注意力+数据+分辨率三板斧重塑 VLM 视觉基座","**一句话**:小鹏 TuringViT 把视觉 Transformer 的注意力、数据、分辨率三个老毛病一起拆——混合线性注意力 + VISTA-Curation 数据治理 + 原生动态分辨率,仅用 SigLIP2 一成的训练数据跑到 SOTA,1536² 分辨率下推理吞吐跑赢主流开源基线 2-3 倍。\n\n视觉编码器正成 VLM\u002FVLA 时代新战场。LLM 把文本端吃干抹净后,所有多模态系统的第一站都是这块\"视网膜\"——决定端侧能不能跑得动、跨任务能不能统一。\n\n7月21日小鹏发布 TuringViT(arXiv:2606.24253,官网 turingvit.github.io),把视觉编码器从\"堆数据的暴力美学\"拽回\"按效果算账的工程经济学\"。\n\n**三板斧**\n\n**Turing 线性注意力**:标准 softmax 注意力随视觉 token 二次增长,高分辨率+多视角视频帧顶到成本天花板。TuringViT 提\"5 层线性注意力+1 层多头注意力\"的混合 Turing Block,让线性注意力负责主要全局聚合,周期性插入少量 MHA 层保住 token 级交互精度,配序列长度感知归一化和输入依赖门控避免细节平滑。\n\n**VISTA-Curation 数据治理**:主流 ViT 靠百亿级图文对,但数据噪声大、对齐度差,堆量收益边际递减。TuringViT 不卷数量、改卷单样本监督价值——多模型多提示词生成候选字幕+统一对比池打分三步筛选。结果只用 0.85B 图文对(SigLIP2-L 训练数据 10%),在 ImageNet-1K 等六项零样本基准拿到 83.6% 平均准确率,超过 10B 数据训练的开源基线。\n\n**原生动态分辨率训练**:MIM 蒸馏初始化→受限动态分辨率(长边 256-512)→原生分辨率放开→图文视频混合训练,配合 2D RoPE,不同尺寸输入无需额外插值。\n\n**关键数字**:1536² 分辨率下 TuringViT-18L 推理吞吐是 Seed1.5-ViT 的 3.04 倍、相比 SigLIP2-ViT-L 提升 2.16 倍——车机、机器人嵌入式算力上能直接吃到的延迟红利。\n\n视觉编码器已经走到 LLM 2026 年的同一个拐点:堆数据边际收益肉眼可见衰减,真正的杠杆是注意力机制的二次方破解+数据治理量级提升+原生多分辨率预训练。TuringViT 不是终点,但它代表视觉编码器从\"scaling 竞赛\"切到\"结构+数据+训练组合拳\"的方向。","https:\u002F\u002Fmp.weixin.qq.com\u002Fs\u002F8SQvN4sH2BR9we2BE7L2zg","4dbd140d-adfd-4c4a-9f9e-50fcfe067281",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":18,"name":19,"slug":19,"description":13,"color":13},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",{"id":21,"name":22,"slug":22,"description":13,"color":13},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"c7319fe1-567a-4522-9d9f-a6faf215ca40","en","XPeng TuringViT cuts vision training cost to one-tenth","**In one sentence**: XPeng's TuringViT tackles all three long-standing problems of vision Transformers — attention, data, and resolution — at once. Hybrid linear attention + VISTA-Curation data governance + native dynamic resolution, using just one-tenth of SigLIP2's training data to reach SOTA, with inference throughput at 1536² resolution running 2–3x faster than mainstream open-source baselines. The vision encoder is becoming the new battlefield in the VLM\u002FVLA era. After LLMs ate the text side clean, the first stop for every multimodal system is this \"retina\" — it determines whether on-device can run and whether cross-task can unify. On July 21, XPeng released TuringViT (arXiv:2606.24253, site turingvit.github.io), pulling the vision encoder from \"brute-force scaling\" back to \"accountable engineering economics\". **The three axes** **Turing linear attention**: standard softmax attention scales quadratically with vision tokens, hitting a cost ceiling at high resolution and multi-view video frames. TuringViT proposes a \"5 linear-attention layers + 1 multi-head attention layer\" hybrid Turing block, letting linear attention handle the main global aggregation, periodically inserting a small number of MHA layers to preserve token-level interaction precision, paired with sequence-length-aware normalization and input-dependent gating to avoid detail smoothing. **VISTA-Curation data governance**: mainstream ViTs rely on tens of billions of image-text pairs, but data noise is high, alignment is poor, and the marginal return from piling on quantity diminishes. TuringViT doesn't race on quantity, it races on per-sample supervision value — multi-model, multi-prompt candidate caption generation + unified contrastive-pool scoring in three steps of filtering. The result: using only 0.85B image-text pairs (10% of SigLIP2-L's training data), it hits 83.6% average accuracy on six zero-shot benchmarks including ImageNet-1K, beating open-source baselines trained on 10B data. **Native dynamic resolution training**: MIM-distillation initialization → constrained dynamic resolution (long edge 256–512) → native resolution released → image-text-video mixed training, paired with 2D RoPE — different-sized inputs need no extra interpolation. **Key numbers**: at 1536² resolution, TuringViT-18L's inference throughput is 3.04x that of Seed1.5-ViT, 2.16x that of SigLIP2-ViT-L — a latency dividend directly available to vehicle and robot embedded compute. The vision encoder has reached the same inflection point LLMs did in 2026: the marginal return from piling on data is visibly decaying, and the real leverage is breaking attention's quadratic blowup + an order-of-magnitude leap in data governance + native multi-resolution pretraining. TuringViT isn't the end, but it represents the direction of vision encoders moving from \"scaling races\" to \"structure + data + training combo punches\".","xpeng-turingvit","2026-07-22T12:00:00Z","2026-07-22T16:08:36.513244Z","2026-08-19T02:08:40.142862Z",true,"agent",159,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"a6fff402-6dee-4dbe-9fcf-ee531b340b12","智平方 NeuroVLA：把\"皮层-小脑-脊髓\"塞进 VLA，端侧机器人反应速度跑到 20 毫秒","zhipingfang-neurovla-cortex-cerebellum-20ms","2026-06-15T12:10:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"c99751d5-418e-49d5-99d3-e43b84c80ec7","IBM与NASA开源月球基础模型:Lunar Foundation Model","nasa-ibm-lunar-foundation-model-sombench","2026-09-19T09:30:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"b9898ba4-65f5-4d80-b9ab-338b01fbd685","NASA 与 IBM 开源月球模型:极区找冰误差降 22%","nasa-ibm-lunar-foundation-model","2026-09-18T13:10:34+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"a18ac6b0-2c9b-4172-87bd-0efe079edc7d","StepAudio 3 Gen：一个模型生成整个声场，官方竞技场两榜居首","stepaudio-3-gen-rvq-autoregressive","2026-09-14T23:07:04+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"0f466258-290e-4c6b-b471-2169ba6a393f","GPT-Live-1 进 API:全双工语音层 0.05 美元一分钟,推理外包给 GPT-6 Astra","gpt-live-1-api-launch","2026-09-12T17:05:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"6062d551-9068-4a9f-ae8e-4e99269cd838","Muse Voice Transcribe 发布:流式转写、20+ 说话人分离、端点检测,Meta 全塞进一个模型","meta-muse-voice-transcribe-streaming-asr","2026-09-05T13:11:00+00:00"]