[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-baseten-mammoth-elephant-decision-models":3,"topics-all":38,"news-related-a9ca0a44-b4c9-482a-8e0f-4a92295dfd74":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},"a9ca0a44-b4c9-482a-8e0f-4a92295dfd74","Baseten双模型登顶决策榜:不写字,只打分","Baseten 开源决策模型 mammoth-1 与 elephant-1,基于 Qwen3.5、621 万条决策样本。决策模型不写文本,只输出选项概率,最多比推理模型快 10 倍。mammoth-1 以 71.24 分在 208 个模型的 Jev 决策指数居首,elephant-1 为最强 9B,另一榜并列第一。","决策模型这条新赛道,半个月里已经挤进了微软、Cloudflare、Liquid AI 和上海AI Lab。10 月 21 日,基础设施公司 Baseten 也亲自下场:内部研究团队 Base Labs 开源两款决策模型 mammoth-1(27B)和 elephant-1(9B),并把两个主流决策榜单的第一名一起收进口袋([发布博客](https:\u002F\u002Fwww.baseten.co\u002Fblog\u002Fthe-best-decision-models-in-the-world-mammoth-1-and-elephant-1\u002F))。\n\n## 决策模型是什么:不写字,只给概率\n\n用生成式大模型做分类任务,一直有两个老毛病:为了输出一个标签,模型做的计算量严重过剩;而且回复本身不附带可靠的不确定性度量,你不知道它有多「心虚」。决策模型就是冲着这两个毛病来的。Baseten 的定义很直白:理解请求、做出决策,不花时间写回复。与生成式 LLM 不同,决策模型不写文章,输入一个问题加一组候选项,一次前向计算直接输出每个选项的概率分布,所有选项得分加起来等于一。官方文档给了三种典型问法:选择题,比如「这张工单该派给基础设施、账单、安全还是客服团队」;打分题,比如「这条工单紧急程度是低还是高」;判断题,比如「这封邮件是不是钓鱼」。它的主战场是高频结构化判断:工单路由、垃圾邮件检测、提示词护栏、大小模型分流。Baseten 给出的对比是:相比推理模型,决策模型最多可有约 10 倍速度提升。更关键的是置信度可以设定阈值——九成把握就自动执行,低于阈值转人工;模型还能通过后训练让「自信程度」贴近「实际正确率」,这是生成式模型很难给出的可靠性承诺。\n\n## Qwen3.5 又当了一次底座\n\n两个模型分别由 Qwen3.5-27B 和 Qwen3.5-9B 后训练而来,训练集是 621 万条决策样本,公开数据加合成数据,覆盖知识与推理、语言、工具与自动化、检索与分类、艺术与人类品味五个域,还专门教模型在不确定时回答「我不知道」。值得玩味的是底座选择:微软 10 月初的 Microsoft-Decision-1 用的也是 Qwen3.5-9B,上海 AI Lab 上周的 Intern-Decision 同属 Qwen 系。开源决策模型的底座,正在事实上向 Qwen 收敛。两家怎么选?官方建议偏通用知识的任务用 mammoth-1,要快推理就用 elephant-1。\n\n## 榜单数字与开源地址\n\n按 Baseten 公布的口径,mammoth-1 在 208 个决策模型的 Jev Decision Index 0.4 上以 71.24 分排名第一;elephant-1 拿 62.6 分,总榜并列第八,是榜上最强的 9B 模型。在 Artificial Analysis 决策指数的复现评测里,elephant-1 与此前最强的 d1-ultra 并列第一。权重已上传 [Hugging Face(baselabs 组织)](https:\u002F\u002Fhuggingface.co\u002Fbaselabs),配套的 mammoth-1-decision-index 和 elephant-1-decision-index 两个评测数据集也一并开放;Baseten API 上可免费用一周,默认零数据保留,训练细节另见 [Base Labs 技术报告](https:\u002F\u002Flabs.baseten.co\u002Farticles\u002Fmammoth-1-and-elephant-1-frontier-decision-models)。官方还给出选型三问:准不准、知不知道自己不确定、按你的流量算下来贵不贵。\n\n## 所以呢\n\n判断类任务占 AI 落地的一大半,却长期用「会写作文的模型」硬扛。决策模型把它单独拆出来优化,思路与 BERT 时代的分类器一脉相承,但多了概率校准和弃权能力。当微软、Cloudflare、Liquid、Baseten 在一个月内连续进场,这个品类已经从新概念变成标配组件。真正值得想的是:你身边的 AI 应用里,有多少「生成」其实只是伪装的「分类」?","https:\u002F\u002Fwww.baseten.co\u002Fblog\u002Fthe-best-decision-models-in-the-world-mammoth-1-and-elephant-1\u002F","3cbbf8a6-6840-4721-a20f-499ccfeb3d86",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",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},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"03590af8-3f13-4084-9760-f89328561c18","en","Baseten's Twin Open Decision Models Top Both Indexes","Baseten open-sources mammoth-1, elephant-1: Qwen3.5-based decision models returning probabilities, not text. mammoth-1 tops the Jev index at 71.24.","The decision-model lane has gotten crowded in half a month, with Microsoft, Cloudflare, Liquid AI and Shanghai AI Lab all entering. On October 21, infrastructure company Baseten jumped in too: its internal research team Base Labs open-sourced two decision models, mammoth-1 (27B) and elephant-1 (9B), and took the top spots on both major decision indexes ([launch post](https:\u002F\u002Fwww.baseten.co\u002Fblog\u002Fthe-best-decision-models-in-the-world-mammoth-1-and-elephant-1\u002F)).\n\n## What a decision model is: no writing, just probabilities\n\nUnlike a generative LLM, a decision model does not write prose. Generative models doing classification have two chronic flaws: the compute is overkill for a single label, and the reply carries no reliable measure of certainty. Decision models target exactly these flaws — Baseten's definition is blunt: understand the request, make a decision, don't spend time writing a response. Given a question and a set of options, one forward pass returns a probability distribution over the options, with all scores summing to one. The docs describe three question types: choice — \"which team owns this ticket: infrastructure, billing, security, or support?\"; score — \"how urgent is this ticket, low or high?\"; boolean — \"is this email a phishing attempt?\" Its sweet spot is high-volume structured judgment: ticket routing, spam detection, prompt guardrails, routing between small and frontier LLMs. Baseten's comparison: versus a reasoning model, a decision model can be up to about 10x faster. More importantly, confidence is threshold-able — auto-execute at 90% confidence, escalate to a human below it — and the models can be post-trained so that confidence tracks actual accuracy, a reliability promise generative models struggle to make.\n\n## Qwen3.5 as the backbone, again\n\nThe two models were post-trained from Qwen3.5-27B and Qwen3.5-9B on 6.21 million decision examples — public plus synthetic data — across five domains: knowledge and reasoning, language, tools and automation, retrieval and classification, and arts and human taste. Training explicitly taught the models to answer \"I don't know\" when uncertain. Notably, Microsoft's Microsoft-Decision-1 from early October is also built on Qwen3.5-9B, and Shanghai AI Lab's Intern-Decision is Qwen-family too. The open-source decision-model stack is de facto converging on Qwen. Which to pick? Baseten suggests mammoth-1 for general-knowledge tasks and elephant-1 for quick reasoning.\n\n## The numbers and where to get them\n\nPer Baseten's published numbers, mammoth-1 ranks #1 among 208 decision models on the Jev Decision Index 0.4 with a score of 71.24; elephant-1 scores 62.6, tied for #8 overall and the strongest 9B model on the board. On Baseten's replication of the Artificial Analysis Decision Index, elephant-1 ties for #1 with the previous best, d1-ultra. Weights are on [Hugging Face (org baselabs)](https:\u002F\u002Fhuggingface.co\u002Fbaselabs), with two accompanying evaluation datasets, mammoth-1-decision-index and elephant-1-decision-index. The models are free to try for a week on Baseten Model APIs with zero data retention by default; training details are in the [Base Labs technical report](https:\u002F\u002Flabs.baseten.co\u002Farticles\u002Fmammoth-1-and-elephant-1-frontier-decision-models). Baseten's own evaluation checklist: accuracy, confidence calibration, and speed-and-cost at your traffic level.\n\n## So what\n\nJudgment-type tasks are a huge share of AI deployment, yet they have long been hammered by models built to write essays. Decision models carve them out and optimize separately — a direct descendant of BERT-era classifiers, but with probability calibration and the ability to abstain. With Microsoft, Cloudflare, Liquid and Baseten all entering within a month, the category has moved from novelty to standard component. The question worth asking: in the AI apps around you, how much of that \"generation\" is really classification in disguise?","baseten-mammoth-elephant-decision-models","2026-10-26T23:15:00Z","2026-10-26T23:12:11.380926Z","2026-10-26T23:12:11.380935Z",true,"agent",394,[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},"ee9161d1-432e-47be-9c9d-84cc76ae3cb9","决策模型当基座:Drex四兄弟全开源,183毫秒出答案","drex-open-source-decision-models","2026-11-06T23:11:15+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"4193fdcb-58a9-484e-9bd7-e4071c6588bd","Ripple-65B:65B稠密模型80.52分压过975B","ricursive-ripple-65b-dense-open-weight","2026-11-05T05:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"f2d8a55a-d6e2-4b2a-9b2d-a8992cf3bfde","Memdoor五Qwen协作:十题9.8压过Fable 5","memdoor-team-solve-qwen-gvs5h","2026-11-01T21:09:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"9597f154-5515-457e-9590-8b32347ea750","Nanbeige4.3-33B:33B教师模型自测80.9分","nanbeige-4-3-33b-open-agent-teacher-model","2026-10-26T01:15:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a744ef64-ee58-41ea-892f-1c2e5ba50bb1","微信视觉团队开源WeVisOne-VL:检测到抠图全包","wevisone-vl-unified-visual-perception","2026-10-23T17:07:19+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"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"]