[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nvidia-dory-1-0-4b-looped-llm":3,"topics-all":38,"news-related-2ddb354a-06f7-4781-b2d5-bcfd1faf38a5":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},"2ddb354a-06f7-4781-b2d5-bcfd1faf38a5","Dory 1.0:英伟达4B循环模型,指令遵循反超9B","英伟达旗下Nemotron Labs发布Dory 1.0:4B循环架构语言模型,同一组层块循环1至4次换精度,上下文256K,OpenMDW-1.1许可开放下载。官方自报评测,IFBench 79分反超Qwen3.5-9B的59,但知识与难题基准落后,vLLM与llama.cpp支持仍在PR阶段。","10 月 16 日，NVIDIA 旗下 Nemotron Labs 低调上线新模型 Dory 1.0-4B。它没有发布会，却选了一条少有人走的路——**循环（looped）语言模型**：参数固定在 4B，\"思考深度\"是推理时可现场调节的旋钮。\n\n## 循环架构：把思考预算做成旋钮\n\n据[模型卡](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FNVIDIA-Nemotron-Labs-Dory1.0-4B-BF16)描述，Dory 是 looped（latent-recurrent）的 normalized-GPT 架构，注意力层采用滑动窗口（SWA）与全注意力交错排布。核心机制是**反复应用同一组共享层块**——官方解释是，这样可以在每个 token 上花更多计算，以此扩展智能。推理时可选 1、2 或 4 次循环：循环越多，单 token 计算越重、精度越高，代价是延迟与算力。\n\n官方推荐默认 2 次循环，并明确大部分精度收益来自第二次循环，之后边际递减。这相当于一份权重自带三挡\"思考预算\"，快与准不再是二选一。\n\n配套规格：上下文最长 256K token，BF16 精度发布，仅支持英文，许可证 OpenMDW-1.1，商用非商用都放行。训练沿用 Nemotron 数据集；后训练管线还包括推理轨迹病态重复过滤，以及对教师模型偶尔产出民族主义叙事的定向清洗——这类细节在模型卡里并不多见。\n\n## 自报跑分：赢在指令遵循，输在知识\n\n官方评测表把 Dory 的 1\u002F2\u002F4 循环与 Qwen3.5-4B、Qwen3.5-9B、Gemma 4 26B-A4B、Qwen3.6-35B-A3B 同表对照。先说明：这些数字全部是 NVIDIA 自报，暂无第三方复现。\n\n赢面集中在指令遵循与 Agent 任务：IFBench 上 Dory 4 循环拿到 79 分，Qwen3.5-9B 只有 59，差距肉眼可见；τ³-bench（Banking）14.69 分是全表最高，把 9B 的 7.22 甩开一倍；长上下文 AA-LCR 的 71\u002F73 也压过所有对照。代码侧，SciCode 35.11 高于 9B 的 28.21，SWE-bench Verified 61.8 对 61 基本追平。\n\n但另一面的数字同样刺眼：GPQA Diamond 72.41 对 9B 的 80.93，HLE 10.89 对 15.43，知识问答 AA-Omniscience 只有 10.08 对 18.05，Terminal-Bench 2.1 的 22.2 也落后于 9B 的 24。规律很清晰——**循环加的是推理算力，补不了参数里的知识**。参数量决定\"知道多少\"，循环次数只决定\"想多深\"。\n\n## 生态位：笔记本能跑，但工具链还没合入\n\nGGUF 量化版（Q5_K_M）约 3.4GB，模型卡直接给了 Apple Silicon 上 llama.cpp 的构建教程，聚合站 The Open Weights 也归入\"笔记本可跑\"一档。DGX Spark 上官方实测 MXFP8 量化对 BF16 提速 1.69-1.84 倍，NVFP4 提速 2.17-2.44 倍。\n\n但尝鲜有门槛：vLLM 支持还在 PR#60149 评审中，llama.cpp 要走 PR#30306 源码构建，都未进正式版。现在上车的，都是愿意自己编译的工具链玩家。\n\n## 所以呢\n\nDory 的真正看点不是某一列跑分，而是 NVIDIA 用一个 4B 模型给\"小参数 + 测试时可调算力\"这条路线投了一票。如果循环架构被验证，边缘设备的模型选择逻辑可能从\"选多大的模型\"变成\"选几挡循环\"。对普通开发者，等 vLLM 合入再上车不迟；对研究者，知识密度与循环深度之间那条边界线，值得盯住。","https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FNVIDIA-Nemotron-Labs-Dory1.0-4B-BF16","474eef8c-e0c3-46cf-adee-c089558220f9",[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},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia",{"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},"f2ed8001-8739-43e4-b0d3-011a3247c342","en","Dory 1.0: NVIDIA's 4B Looped LLM Outscores 9B on IFBench","NVIDIA ships Dory 1.0, a 4B looped LLM trading compute for accuracy; self-reported IFBench 79 tops Qwen3.5-9B's 59, knowledge benches lag.","On October 16, NVIDIA's Nemotron Labs quietly published a new entry: Dory 1.0-4B. There was no flagship launch event and no flood of press releases, but the architecture takes a road few have traveled: a **looped language model**, where parameters stay fixed at 4B while \"thinking depth\" becomes a dial you can turn at inference time.\n\n## A looped architecture: thinking budget as a knob\n\nAccording to the [model card](https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FNVIDIA-Nemotron-Labs-Dory1.0-4B-BF16), Dory is a looped (latent-recurrent) normalized-GPT whose attention layers interleave sliding-window attention (SWA) with full attention. The core mechanism is **repeatedly applying a shared block of layers** — in NVIDIA's words, the model can spend more compute per token to scale intelligence. At inference you choose 1, 2, or 4 loops: more loops mean more compute per token and higher accuracy, at the cost of latency.\n\nThe card recommends two loops as the default and notes that most of the accuracy gain comes from the second loop, with diminishing returns beyond. For deployers, that's a 4B model shipping with three \"thinking budget\" presets — speed and accuracy stop being an either-or choice over a single set of weights.\n\nSupporting specs: up to 256K tokens of context, released in BF16, English-only, OpenMDW-1.1 license permitting commercial and non-commercial use. Training draws on NVIDIA's Nemotron datasets; the post-training pipeline includes filtering pathological repetition in reasoning traces and targeted removal of trajectories where teacher models occasionally produced nationalist narratives — a detail you rarely see spelled out in a model card.\n\n## Self-reported benchmarks: wins instruction following, loses knowledge\n\nThe official table pits Dory at 1\u002F2\u002F4 loops against Qwen3.5-4B, Qwen3.5-9B, Gemma 4 26B-A4B, and Qwen3.6-35B-A3B. Caveat first: all numbers are NVIDIA's own, with no third-party replication yet.\n\nThe wins cluster in instruction following and agentic tasks: IFBench 79 at four loops versus 59 for Qwen3.5-9B (53.83 for the 4B); τ³-bench Banking at 14.69 — the highest in the row, double the 9B's 7.22; long-context AA-LCR at 71\u002F73 also beats every comparator. On code, SciCode 35.11 clears the 9B's 28.21, and SWE-bench Verified 61.8 essentially matches 61.\n\nThe other side is equally blunt: GPQA Diamond 72.41 versus 80.93 for the 9B, HLE 10.89 versus 15.43, AA-Omniscience at 10.08 versus 18.05, and Terminal-Bench 2.1 at 22.2 behind the 9B's 24. The pattern is clear — **looping adds inference compute; it cannot add the knowledge stored in parameters**. A 4B parameter count decides how much the model knows; loop count only decides how deeply it thinks. That draws a crisp boundary around looped architectures.\n\n## Ecosystem niche: laptop-runnable, but the toolchain isn't merged\n\nThe GGUF quantized export (Q5_K_M) is about 3.4 GB, the model card ships a llama.cpp build guide for Apple Silicon, and aggregator The Open Weights files it under \"runs on a laptop.\" On DGX Spark, NVIDIA reports MXFP8 quantization running 1.69-1.84x faster than BF16 and NVFP4 at 2.17-2.44x.\n\nEarly adoption has a catch: vLLM support is still under review (vllm-project#60149), and llama.cpp requires building from PR #30306 — neither has landed in an official release. Everyone boarding now is a toolchain player willing to compile.\n\n## So what\n\nDory's real story isn't any single benchmark column; it's NVIDIA casting a vote for \"small parameters plus tunable test-time compute\" with a 4B model. If looped architectures validate, model selection on edge devices may shift from \"how big a model\" to \"how many loops.\" For most developers, waiting for the vLLM merge is reasonable; for researchers, the boundary between knowledge density and loop depth is the thing to watch.","nvidia-dory-1-0-4b-looped-llm","2026-10-20T21:09:38Z","2026-10-20T21:09:43.570268Z","2026-10-20T21:09:43.570283Z",true,"agent",58,[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},"b86d15f9-d4de-4013-ab3c-456717759a47","NVIDIA Audex 把\"听懂+听懂+说出\"压进一个 30B MoE：单 Transformer 解码器统一文本与音频推理","nvidia-audex-30b-moe","2026-07-08T10:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"5736961c-9502-40d0-a39b-27f8ceb4380a","AI造AI:Recursive开源笔记本编程模型Flint","recursive-flint-laptop-coding-model","2026-10-22T19:05:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"faf5c85e-ed83-432a-8b55-80c9356862c5","腾讯优图开源5B全模态解析:文档榜96.96超GPT-5.2","youtu-parsing-omni-5b","2026-10-19T19:09:27+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"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"]