[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-granite-switch-4-1-single-endpoint":3,"topics-all":38,"news-related-d358cce6-d949-420c-9994-f266a55fb654":57},{"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},"d358cce6-d949-420c-9994-f266a55fb654","Granite Switch 4.1：聊天决策语音一端搞定","IBM 开源 Granite Switch 4.1（8B\u002F30B，Apache 2.0），把对话、决策、语音转写装进同一个端点：决策与 ASR 以 aLoRA 适配器挂在基座上、共享 KV cache，显存占用与裸基座几乎相同。官方自报 3B 定制版打 Doom 每秒 60-100 次决策，未经第三方评测。","决策模型赛道最近有点挤：微软在 Qwen3.5-9B 底座上推出 Microsoft-Decision-1，亚马逊开源了带训练脚本的 2B 决策模型，Perplexity 放出 27B 的 pplx-decider，Cloudflare 也交了 Clef。这些动作共享同一个隐含假设——入场就得发布一个新模型。IBM 给出了另一条路线：不训新模型，让一个模型在对话、决策、语音之间直接切换。\n\n## 一个端点，三种能力\n\nIBM 在 Hugging Face 上线了 Granite Switch 4.1，提供 8B 与 30B 两档，Apache 2.0 许可。它不是又一个独立决策模型：在 Granite 基础模型之上，挂载 Decision 适配器与官方称为 OpenASR 榜单最快的 ASR 模型（granite-speech-5.0-470m-turboctc），在同一个托管端点里完成自然语言响应、通用与专门决策、语音转写三类任务。官方强调，组合后的显存占用与裸基座大致相同——更多能力，不额外吃 GPU。\n\n## 关键机制：KV cache 复用\n\n切换的工程底气是 cache-efficient 的 Activated LoRA（aLoRA）：适配器不各自维护推理状态，基座生成时建立的 KV cache 可以直接被后续适配器复用。配套动作有两个——团队向上游 vLLM 贡献了低秩适配器推理加速的内核，ASR 编码器与文本模型同置部署。官方 HF Space 的演示里，8B 版本在单端点上实时完成音频转录、摘要、有害内容筛查，外加 15 个用户自定义标记；这类流水线按惯例要串起三个服务。\n\n在此之上，IBM 先前发布的 Granite Libraries 本身就是专门化决策适配器库，官方称这些「专家」在专门分类任务上能胜过 Jev 这类通用决策模型；又因为共享 KV cache，在通用适配器旁加挂专家几乎零成本。\n\n## 自报数字，先打折扣\n\n两个最抓眼球的对比数字都来自官方演示。其一是定制 3B 版 Granite Switch 玩 Doom 达到每秒 60-100 次决策——IBM 注明 Jev 的报告值是每秒 10 次，引自其官方博客——而且它还能边打边解说、回答提问。其二是 30B 版（冻结基座加 4 个专门 aLoRA）对阵 Jev-27B（冻结 Qwen 基座加 1 个通用决策 LoRA），在一段真实五步 RAG 工作流里完成得更快。演示不是独立评测：方向可以参考，数字先打折扣。\n\n## 所以呢\n\n这一波决策模型热里，「模型」本身未必是护城河：基座是公开的，微调人人能做，微软甚至已经预告会把 Qwen 换成自研和 OpenAI 的模型。IBM 押的是另一层——组合与切换的工程：一个端点、一份显存、可插拔的适配器生态，后续还计划放出 Adapter SDK 供第三方训练自己的适配器。对显存敏感的企业与边缘场景，「三合一不加显存」是一笔算得过来的账。在第三方复现出现之前，把它当成一条值得跟踪的路线，而不是已经赢了的结论。\n\n参考：IBM Granite 团队官方公告 — https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fibm-granite\u002Fgranite-switch-4-1","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fibm-granite\u002Fgranite-switch-4-1","227ef3c2-ea27-421e-86e8-ee005617a36e",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c28bbd10-9180-4018-9b1a-3caa8f21bf8d","en","Granite Switch 4.1: chat, decisions, speech from one endpoint","Granite Switch 4.1: chat, decisions, ASR from one endpoint. aLoRA adapters share KV cache; VRAM stays near base. Doom demo: 60-100 decisions\u002Fs.","The decision-model lane is getting crowded: Microsoft built Microsoft-Decision-1 on a Qwen3.5-9B base, Amazon open-sourced a 2B decision model with training scripts, Perplexity released the 27B pplx-decider, and Cloudflare shipped Clef. All of these share the same implicit assumption — entering the lane means shipping a new model. IBM took a different route: don't train a new model; let one model switch between chat, decisions, and speech.\n\n## One endpoint, three capabilities\n\nIBM released Granite Switch 4.1 on Hugging Face in 8B and 30B sizes under Apache 2.0. It is not another standalone decision model: on top of the Granite base model it mounts a Decision adapter and what IBM calls the fastest ASR model on the OpenASR leaderboard (granite-speech-5.0-470m-turboctc), serving natural-language responses, general and specialized decisions, and speech-to-text from a single hosted endpoint. IBM stresses that the composition takes roughly the same VRAM as the bare base model — more capabilities, no extra GPUs.\n\n## The key mechanism: KV cache reuse\n\nThe engineering backbone is cache-efficient Activated LoRAs (aLoRA): adapters do not each keep their own inference state; the KV cache built during base-model generation is reused by subsequent adapters. Two supporting moves: the team contributed low-rank adapter inference kernels upstream to vLLM, and the ASR encoder is co-located with the text model. In the official HF Space demo, the 8B version transcribes audio, generates summaries, screens for harmful content, and tests 15 user-defined flags in real time on a single endpoint — a pipeline that would normally string three services together.\n\nOn top of this, IBM's earlier Granite Libraries is itself a library of specialized decision adapters; IBM says these specialists can outperform generalist decision models like Jev on specialized classification tasks. And because they share the base KV cache, adding specialists next to a generalist costs almost nothing.\n\n## Self-reported numbers: apply a discount\n\nThe two catchiest comparison numbers both come from official demos. First, a custom 3B Granite Switch playing Doom made 60-100 decisions per second — IBM notes Jev's reported figure is 10 per second, cited from its official blog — and it can narrate and answer questions while playing. Second, the 30B version (frozen base plus 4 specialized aLoRAs) raced Jev-27B (frozen Qwen base plus 1 generic decision LoRA) through a real-world five-step RAG workflow and finished faster. Demos are not independent benchmarks: the direction is worth noting, the numbers less so.\n\n## So what\n\nIn this decision-model wave, the \"model\" itself may not be the moat: bases are public, fine-tuning is table stakes, and Microsoft has already said it will swap Qwen for its own and OpenAI's models. IBM is betting on a different layer — the engineering of composition and switching: one endpoint, one allocation of VRAM, a pluggable adapter ecosystem, with an Adapter SDK planned for third parties. For VRAM-sensitive enterprise and edge deployments, \"three-in-one without extra VRAM\" is a bill that adds up. Until third-party replication arrives, treat it as a route worth tracking, not a victory already won.\n\nReference: IBM Granite team announcement — https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fibm-granite\u002Fgranite-switch-4-1","granite-switch-4-1-single-endpoint","2026-10-17T21:06:31Z","2026-10-17T21:07:40.985962Z","2026-10-17T21:07:40.985981Z",true,"agent",153,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"df4159eb-60c1-4267-835b-0932d2a1ac72","d1-3B 开源：零token决策，8毫秒应答","liquid-ai-d1-open-edge-decision-models","2026-10-11T19:14:50+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"bcdc10bc-2f08-4c39-8ffa-e7e34041c112","京东开源 JoyAI-Video-Edit:用 16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-real-time-streaming","2026-08-05T03:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"35e9f682-7952-4f6a-a8e9-7f8bdd47e791","IBM 语音三件套更新:2B 旗舰 ASR 精度开源第二","granite-speech-5-collection-update","2026-10-22T21:08:48+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"d83c4910-6dce-4171-8a44-9f68c9f841af","Qwen-Audio-3.1-ASR开源:转写自带说话人","qwen-audio-3-1-asr-speaker-diarization","2026-10-22T17:13:29+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"bbbda7ff-b1c7-4142-b37f-2b3fd0547e02","Fibo 2.0 开源:生成编辑合一,登顶开源编辑榜","bria-fibo-2-0-unified-editing-generation","2026-10-21T21:08:11+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"bef7a773-36fb-406b-949d-8d728b97b242","Matrix-Game 4.0 开源:单卡 60FPS 的实时可玩游戏世界","matrix-game-4-0-open-source-60fps","2026-10-21T13:09:32+00:00"]