[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mobile-ai-factory-on-device-slm":3,"topics-all":38,"news-related-1c9b3e02-69f2-452f-9137-ff9c2ddcf7fb":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"1c9b3e02-69f2-452f-9137-ff9c2ddcf7fb","手机上的模型工厂:小模型自动定制,五项反超前沿模型","Gollakota 团队发布 Mobile AI Factory:输入任务与手机,Agent 自动选数据、调超参,分桶搜索 27 种基模与量化组合,逐台实测手机速度与内存。10 项任务里 5 项超过 Claude Sonnet 5 与 DeepSeek v4 Flash 五样本成绩,最高领先 0.342。","同一个模型,换台手机就像换了个性格。论文实测:llama.cpp 的 GPU prefill 在 OnePlus 15 上比 CPU 快 5.4 倍,到 OnePlus 8 Pro 上反而慢 7.4 倍;Qwen3-4B 的 8-bit 版本在 OnePlus 15 上每秒解码 12.6-13.8 个 token,换成 OnePlus 8 Pro 就掉到 0.7-4.8。端侧小模型真正的难题从来不是「能不能跑」,而是每台手机都得重新调一遍——Facebook 一款 app 就要适配超过一万种机型,开源社区每周还冒出大约两个新基模,靠专家手工逐台调优,只有最大的公司养得起这种团队。\n\n## 工厂怎么运转\n\nGollakota 团队的方案叫 Mobile AI Factory:输入任务规格和目标手机,产出一个微调加量化、刚好塞进这台手机的小模型。核心是一个循环:每轮由编排 LLM(Claude Sonnet 5)从三种干预里挑一个——调超参(LoRA rank、学习率、训练轮数等)、去任务源里挖更多真实数据、或者让本地托管的教师模型(Qwen3.6-35B-A3B)针对最拖准确率后腿的失败类别生成合成数据。固定规则负责兜底:凡把准确率调低的改动一律回滚到当前最优,训练前还有污染检查,把混进保留测试集的行剔掉。\n\n搜索空间按体积分桶:9 个基模(SmolLM2、Gemma 3、Qwen3、Qwen3.5,参数量 135M 到 4B)乘 4-bit、8-bit、16-bit 三档精度,共 27 种组合,按磁盘体积切成 5 个桶,从最小的桶往上搜。准确率目标也是自适应的:初始取教师模型的五样本成绩、下限 0.8,太容易达标就主动上调继续压榨。最关键的部件是手机画像器:每台手机一次性实测全部组合在 llama.cpp 与 MNN、CPU 与 GPU 下的 prefill\u002Fdecode 速度、峰值内存和能耗,4 台手机共 364 组配置,画像跨任务复用,搜索只在真跑得动的模型里挑。\n\n## 战绩单\n\n10 个任务、4 台 Android 手机(从当前的旗舰机型到 GPU 因缺 OpenCL 用不了的 Redmi 9A):在没见过的测试集上,工厂产出的小模型在 5 项任务上同时超过 Claude Sonnet 5 和 DeepSeek v4 Flash 的五样本成绩,领先 0.024 到 0.342,另有 3 项打平。输出格式越严格,优势越大:生物医学实体抽取任务上,一个 517MB 的模型拿到 0.828,Claude 五样本只有 0.652。\n\n搜索本身也划算:实体抽取任务上,分桶搜索与穷举训练全部 27 种变体只差 3 个点,GPU 时间省 19 倍。Galaxy S23 的资源上限卡掉最优模型时,工厂退而返回的次优模型只差 0.013。首次微调后,搜索还能给最小的模型中位数再加 8 个点,最大的加 2 个点;9 次因目标上调而继续训练的运行里,7 次又涨了 3.6 个点。\n\n## 规格表推理靠不住\n\n对照组更有意思:给 Claude Opus 5.5 完整的硬件与模型规格、但不给实测数据,让它判断某配置能否满足 app 需求——简单和中等场景漏判 25%,困难场景漏判 60%,错就错在运行时内存和该用 CPU 还是 GPU 这种规格表上不写的项。最好的配置本来就随手机变,推理替代不了测量。\n\n## 所以呢\n\n论文贡献声明里,作者自称这是「据我们所知第一个」面向给定任务加给定手机的此类系统——优先权由他们认领,工程事实摆在这里:端侧 AI 的墙不是模型能力,是硬件碎片化,破墙靠的是把「每台手机一个模型」变成自动化流水线。对独立开发者,这套思路比任何单一新模型都值钱。一个提醒:论文称代码开源于 GitHub,但截至发稿该仓库仍无法访问,想复现的还得再等等。\n\n原文:https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.13509","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.13509","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"20e72bc9-6eb8-407e-a8bc-c8e5c0c916c4","en","Mobile AI Factory Beats Frontier Models on Five Phone Tasks","Gollakota's Mobile AI Factory auto-builds fine-tuned, quantized SLMs per task and phone; five of ten tasks beat two five-shot frontier models.","The same model behaves like a different animal on a different phone. In their measurements, llama.cpp GPU prefill is 5.4x faster than CPU prefill on the OnePlus 15, yet 7.4x slower on the OnePlus 8 Pro; Qwen3-4B at 8-bit decodes 12.6-13.8 tokens\u002Fs on the OnePlus 15 but only 0.7-4.8 tokens\u002Fs on the 8 Pro. The hard part of on-device language models was never whether it can run - it is that every phone needs its own tuning pass. The Facebook app alone ships to more than ten thousand distinct phone and tablet models, and roughly two new open foundation models appear every week. Hand-tuning each task on each device is a luxury only the largest companies can staff.\n\n## How the factory runs\n\nThe team's answer, Mobile AI Factory, takes a task specification and a target phone, and produces a fine-tuned, quantized small language model sized for that device. The core is a loop: each iteration, an orchestrator LLM (Claude Sonnet 5) picks one intervention - tune hyperparameters (LoRA rank, learning rate, epochs), mine more real data from the task's sources, or have a locally hosted teacher (Qwen3.6-35B-A3B) generate synthetic examples aimed at the failure categories hurting accuracy most. Fixed rules do the janitorial work: any change that lowers accuracy is rolled back to the best model so far, and a contamination check removes rows that also appear in the held-out test set.\n\nThe search space is bucketed. Nine base models (SmolLM2, Gemma 3, Qwen3, Qwen3.5; 135M to 4B parameters) times three precisions (4-bit, 8-bit, 16-bit) give 27 combinations, grouped into five buckets by on-disk size and searched from smallest up. Accuracy goals adapt too: the initial target is the teacher's five-shot score with a floor of 0.8, and if it is reached too easily the orchestrator raises it and keeps pushing. The crucial component is the phone profiler: once per phone, it measures prefill and decode speed, peak memory, and energy across llama.cpp and MNN runtimes on CPU and GPU - 364 configurations across four phones - and reuses the profile across tasks, so the search only considers models the phone can actually run.\n\n## The scorecard\n\nAcross ten tasks and four Android phones (from a current flagship down to the Redmi 9A, whose GPU the inference engines cannot use for lack of OpenCL): on unseen test sets, the factory's small models beat both Claude Sonnet 5 and DeepSeek v4 Flash given five examples on five tasks, by 0.024 to 0.342, and match them on three more. The stricter the output format, the bigger the gap: on biomedical entity extraction, a 517 MB model scores 0.828 where five-shot Claude manages 0.652.\n\nThe search itself is frugal: on entity extraction, bucketed search lands within 3 points of exhaustively training all 27 variants while using 19x less GPU time. When the Galaxy S23's resource limits rule out the factory's best model for call summaries, it returns an alternative within 0.013. After the first fine-tune, search adds a median of 8 points on the smallest models and 2 on the largest; in seven of the nine runs where a raised goal triggered more training, the model improved by 3.6 points.\n\n## Spec-sheet reasoning fails\n\nThe control experiment is the kicker: given the full hardware and model specifications but no measurements, Claude Opus 5.5 misses the app's requirements in 25% of easy and moderate cases and 60% of hard ones - misjudging runtime memory and the right backend, neither of which appears on a spec sheet. The best configuration changes across phones, and reasoning cannot substitute for measurement.\n\n## So what\n\nIn their contributions list, the authors call this, to their knowledge, the first agentic system that builds a fine-tuned, quantized SLM for a given task and a given phone - priority is theirs to claim; the engineering facts stand on their own. The real wall for on-device AI is hardware fragmentation, not model capability, and the way through is turning one-model-per-phone into an automated pipeline. For independent developers, that idea is worth more than any single new checkpoint. One caveat: the paper says code is available on GitHub, but as of writing the repository still returns 404 - reproduction fans will have to wait.\n\nPaper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.13509","mobile-ai-factory-on-device-slm","2026-11-12T23:15:00Z","2026-11-12T23:08:37.089355Z","2026-11-12T23:08:37.089368Z",true,"agent",185,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"e8f13686-3f2f-468a-b974-55634bb30795","Metal Sandbox:Mac 集群的 GPU 调度术","metal-sandbox-mac-cluster-gpu-scheduling","2026-11-15T03:15:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"76bc2964-2411-42b7-a8b9-389bbc6d417e","SteerDistill:量化Agent失手,老师闭环来救","steerdistill-quantized-agent-teacher-steering","2026-11-12T19:06:48+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"836384aa-b52a-4c94-8845-c6e6610def08","OpenAI「无APP手机」原型现场演示：Agentic操作系统登场，UI实时生成、推理甩给云端GPT，2027年量产","openai-agentic-os-no-app-phone-2027","2026-06-02T01:06:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"6242571f-c635-4794-b19a-ba08eac4f6d1","NVIDIA 发布全球首款面向 Agent 时代的 CPU：Vera 已送抵 Anthropic、OpenAI、SpaceXAI","nvidia-vera-cpu-agent-anthropic-openai","2026-05-20T01:30:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"648ee35f-321e-4cc7-8e88-9a9d3ebae033","小模型每晚自学:一年后从 52 分掉到 5 分","on-device-llm-self-training-judge-failure","2026-11-19T03:30:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"03403319-8597-4876-8661-fad74ee0ac48","屏幕上的警告,智能体最高92%照做","computer-use-agents-cross-screen-harm","2026-11-18T23:30:00+00:00"]