[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-drex-open-source-decision-models":3,"topics-all":38,"news-related-ee9161d1-432e-47be-9c9d-84cc76ae3cb9":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},"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","Nace.AI 开源 Drex 决策模型家族,8B 到 27B 共四款,把判断题与选择题从文本生成里拆出来:一次前向直接输出各选项概率,不生成文本,状态可以是文档、记录或智能体上下文。在论文对比的决策模型里,Max 得 67.52,胜过 Jev 的 60.70;单次请求 183 毫秒,权重与训练配方全部公开。","生产环境里一大类「大模型任务」,其实既不用写文章也不用写代码:这张工单要不要升级、两条摘要哪条更相关、这段内容是否违规、候选人过不过初筛。过去几年的默认解法,是让生成式模型跑一长串思维链、吐出文本、再写解析——慢、贵,而且同一个问题换个选项顺序,答案可能就变了。Nace.AI 新放出的 Drex 家族换了思路:输入一段上下文和一组选项,一次前向传播直接给出每个选项的概率,不生成任何文本。配套的 51 页技术报告([arXiv:2611.07854](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.07854))把方法论、消融和部署细节一次讲完,四款权重全部开放。\n\n## 决策模型:不是更大的模型,是另一种读出方式\n\n「决策模型」这个词沿袭自 Jev 的 System One 接口:给定一个状态(文档、记录或智能体上下文,可带图片)和若干问题(选择题、判断题、打分题),模型直接返回选项上的概率分布,而不是自然语言。报告认为,判别式读出补的正是自回归链路在生产里的三块短板:答案不稳定、输出要解析、成本和延迟随生成长度膨胀。分类、打分、嵌入这三种判别任务的输出形状在部署时就固定了;决策模型要啃的是第四种——候选集合由用户在推理时临时给定,这恰好是业务里最常见的形态。\n\n## 一个接口,四个底座\n\nDrex 不是单点模型,而是一个家族。DLM 8B 基于 NVIDIA 的 Efficient-DLM-8B;v1.5 9B 基于 MiMo-V2.6-Distill-Qwen-9B;v2.0 是 35B 参数、约 3B 激活的 MoE,基于 Qwen3.6-35B-A3B,10 月 13 日发布时在 Decision Index 0.3.1 公开榜拿到 61.25;旗舰 Max 是 27B 稠密多模态模型,从 Qwen3.8-27B 训练而来,10 月 14 日跟进。四款共享同一套请求格式,架构却分成两派:pointer 系(DLM、v1.5、v2.0)把上下文编码一次、多问共享,成本更低;Max 对每个问题重新编码完整状态,并去掉全注意力层的因果掩码,再用一个无偏置线性层把末层隐状态映射为 255 个选项码的 logits——单题更贵,换更高精度。\n\n## 训练配方:三段式,消融数字都对得上\n\nMax 的训练分三段。监督微调打头,损失里加两项一致性惩罚:选项顺序被打乱、状态里的句子和字段被洗牌,预测分布都不该变。接着是论文主推的 RLAF(基于调整反馈的强化学习):奖励混合「可验证的正确答案」与「裁判模型分给各选项的概率」,再用 KL 散度把策略拴在自己的 SFT 检查点附近。最后微调图像编码器及其投影,让模型直接看图做决策。论文消融给出三个硬数字:RLAF 把 Decision Index 分数抬高 2.0;一致性损失把换序翻转率从 6.3% 压到 4.2%;视觉阶段不损失纯文本任务的准确率。\n\n## 跑分与延迟:官方口径很强,独立榜单更冷静\n\n按论文口径(Decision Index 0.4.1 公开套件,2026 年 10 月 22 日榜单),Max 以 67.52 领先 Jev 1.14.0 的 60.70,在报告对比的决策模型中刷新纪录;在维护方私有测试上得 66.60。报告也没有回避独立评测:Artificial Analysis 独立运行的 Decision Index 上,Max 以 64 分列决策模型第二,领先的三款模型为 65 分。Decision Bench v4 上,它与 d3-turbo、GPT-6 Sol 统计打平,落后 Claude Opus 5.5 约 1.6 分。延迟是这类模型的存在理由:Max 单次典型请求 183 毫秒,对照 GPT-6 Sol 约 2.0 秒、Claude Opus 5.5 约 3.5 秒。\n\n## 开源程度与部署门槛\n\n代码以 Apache-2.0 放出,Max、v2.0、v1.5 的权重走 Nace.AI Open RAIL-M 许可,DLM 沿用底座的 CC BY-NC 4.0。运行侧除 Python 参考实现外,还有 llama.cpp 与 Ollama 的 fork、brew 可装的 Kapitsa 和 MLX:affine Q4 版在 24GB 内存的 Apple Silicon Mac 上就能跑 Max,Q8_0 需要 48GB;Max 上下文上限 262,144 个 token。三款主力模型同时上了托管 API 与 OpenRouter,截至发稿 [GitHub 仓库](https:\u002F\u002Fgithub.com\u002Fnace-ai\u002Fdrex-decision-models) 84 星、57 次提交。论文还披露了推理优化的细节:用 FlashInfer 的 Blackwell prefill 核替换自写 Triton 核,8k 到 128k 状态延迟降 9-16%,98.7% 的首选选项与换核前一致;把全部 128 个 MLP 投影量化到 NVFP4 后,中位延迟再降约 8%,但在 When2Call 上比 BF16 参考高 2.00 分、超出每数据集 1.5 分的自设阈值——团队人工复核后仍然发布,并把差异写进了论文。\n\n## 值得留个心眼的地方\n\n报告自己承认:除 Max 的一项受控消融外,多数比较是关联而非因果,哪些设计真正驱动提升仍待实验;Max 每题都要重编码完整状态,长上下文多问场景成本高于 pointer 系;翻转率降到 4.2% 也仍非零,顺序敏感的生产场景要留冗余。另一面,过去一个月微软、亚马逊、Cloudflare、上海 AI Lab 相继押注决策模型,Drex 则把「家族化、全开源加完整方法论」一次打包放出。当全行业都在给生成式推理堆算力时,不妨数一数:你应用里的「生成」,有多少其实只是「选择」?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.07854","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",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},"94804da9-6911-4d91-b58f-b33f5d5e02c7","en","Drex: Four Open Decision Models, One Forward Pass, 183 ms","Nace.AI open-sources Drex, four decision models returning option probabilities in one forward pass; Max scores 67.52 vs 60.70 for Jev, at 183 ms.","A large share of “LLM tasks” in production write neither prose nor code: should this ticket be escalated, which of two summaries is more relevant, is this content compliant, does this candidate pass screening. The default answer of recent years is a long chain-of-thought generation followed by parsing — slow, expensive, and sometimes the answer flips when the options are merely reordered. The Drex family from Nace.AI takes a different route: pass in a context and a set of options, and a single forward pass returns a probability for every option, with no text generated at all. A 51-page technical report ([arXiv:2611.07854](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.07854)) documents the methodology, ablations and deployment details, and all four sets of weights are open.\n\n## Decision models: a different readout, not a bigger model\n\nThe term follows the System One interface introduced by Jev: given a state (a document, record or agent context, optionally with images) and questions (multiple choice, yes\u002Fno, or a rating scale), the model returns a probability distribution over the options instead of natural language. The report argues that discriminative readouts patch three pain points of autoregressive pipelines in production: inconsistent answers, outputs that must be parsed, and cost and latency that grow with generation length. Classification, scoring and embedding fix their output shape at deployment time; decision models tackle the fourth kind, where the candidate set is supplied by the user at inference time — the shape most business decisions actually take.\n\n## One interface, four backbones\n\nDrex is a family rather than a single model. DLM (8B) builds on NVIDIA Efficient-DLM-8B; v1.5 (9B) builds on MiMo-V2.6-Distill-Qwen-9B; v2.0 is a 35B mixture-of-experts with about 3B active parameters on the Qwen3.6-35B-A3B backbone, scoring 61.25 on the public Decision Index 0.3.1 when it shipped on October 13; the flagship Max is a dense 27B multimodal model trained from Qwen3.8-27B, released October 14. All four share one request format but split into two architectural camps: the pointer models (DLM, v1.5, v2.0) encode the context once and share it across questions, while Max re-encodes the full state for every question, removes the causal attention mask in its full-attention layers, and maps the final hidden state to logits over 255 option codes through a bias-free linear layer — costlier per question, in exchange for accuracy.\n\n## A three-stage recipe, with ablations to match\n\nMax trains in three stages. Stage one is supervised fine-tuning with two consistency penalties: the predicted distribution must not change when the options are permuted, nor when the sentences or fields of the state are shuffled. Stage two is RLAF — reinforcement learning from adjusted feedback — where the reward blends the verified correct answer with the probabilities a judge model assigns to the options, and a KL term keeps the policy close to its own SFT checkpoint. Stage three tunes the image encoder and its projection so the model can decide directly from images. The ablations are concrete: RLAF lifts the Decision Index score by 2.0 points; the consistency terms cut the option-order flip rate from 6.3% to 4.2%; and the vision stage does not reduce accuracy on text-only inputs.\n\n## Scores and latency: strong on the report numbers, sober on the independent index\n\nOn the report numbers (Decision Index 0.4.1 public suite, leaderboard of October 22, 2026), Max scores 67.52 against 60.70 for Jev 1.14.0 — a new best among the decision models compared in the report — and holds 66.60 on maintainer-run private tests. The report also faces the independent view: on the Artificial Analysis Decision Index, Max ranks second among decision models with 64, versus 65 for the three leading models. On Decision Bench v4 it is statistically tied with d3-turbo and GPT-6 Sol and trails Claude Opus 5.5 by 1.6 points. Latency is the reason this model class exists: a typical request to Max completes in 183 ms, against roughly 2.0 s for GPT-6 Sol and 3.5 s for Claude Opus 5.5.\n\n## Openness and deployment\n\nThe code is Apache-2.0; Max, v2.0 and v1.5 weights ship under the Nace.AI Open RAIL-M license, while DLM follows the CC BY-NC 4.0 license of its backbone. Beyond the Python reference server there are llama.cpp and Ollama forks, a Homebrew-installable runner and MLX: the affine-Q4 checkpoint runs Drex Max on Apple Silicon Macs with 24 GB of memory, Q8_0 needs 48 GB, and Max supports contexts up to 262,144 tokens. Three models are also served through a hosted API and OpenRouter; at writing, the [GitHub repository](https:\u002F\u002Fgithub.com\u002Fnace-ai\u002Fdrex-decision-models) sits at 84 stars and 57 commits. The report even discloses its inference surgery: swapping hand-written Triton kernels for the FlashInfer Blackwell prefill kernel cuts latency by 9–16% on 8k–128k-token states, with 98.7% of top options agreeing with the previous kernels; quantizing all 128 MLP projections to NVFP4 shaves roughly 8% off median latency, but it landed 2.00 points above the BF16 reference on When2Call — past the 1.5-point-per-dataset limit the team set for itself. They shipped it after manual review and reported the difference.\n\n## Where to stay skeptical\n\nThe report admits its own limits: outside one controlled ablation on Max, most comparisons are associations rather than causal conclusions, and which design changes actually drive the gains remains open. Max re-encodes the full state for every question, which costs more than the pointer models on long multi-question contexts, and a 4.2% flip rate is not zero, so order-sensitive production use still needs headroom. Meanwhile Microsoft, Amazon, Cloudflare and Shanghai AI Lab have all pushed decision models in the past month, and Drex bundles a family, open weights and a full methodology into one release. While the industry piles compute onto generative reasoning, it is worth counting how much of the “generation” in your application is really just “selection”.","drex-open-source-decision-models","2026-11-06T23:11:15Z","2026-11-06T23:11:18.181039Z","2026-11-06T23:11:18.181046Z",true,"agent",91,[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},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"d056f67b-7e0d-4e44-8d39-e31ea50deeae","Bonsai 2 27B 三元压缩:Qwen3.8 压到 5.9 GB,benchmark 留存 98.2%","bonsai-2-27b-ternary-qwen3-8-compression","2026-09-17T15:47:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"637f84e0-e6dc-490a-bba1-879f6527bdd5","Qwen3.8-Max 2.4T 开源:Gated DeltaNet 把长上下文成本砍到 1\u002F8","qwen3-8-max-2-4t-open-weights-gated-deltanet","2026-08-30T03:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"40095b51-97b0-4fd4-9b1d-f636c970572e","阿里 Qwen 团队发布 Qwen3.8-Max:2.4 万亿参数 MoE 模型首度开放权重","qwen3-8-max-2-4t-moe-open-weights","2026-08-07T02:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"df01dae0-0940-4947-a019-31c57066132c","阿里千问 Qwen3.8 预览版上线:2.4T 参数,把开源旗舰抬到 Fable 5 同一档","qwen-3-8-preview-2-4t","2026-07-19T10:01:00+00:00"]