[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-agnes-3-0-flash-preview-open-weights":3,"topics-all":38,"news-related-550cee5e-18e8-4236-9304-7207ebc221a8":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},"550cee5e-18e8-4236-9304-7207ebc221a8","Agnes 3.0 Flash 开源:72 层仅 18 层带 KV 缓存,33B 单卡跑 262k 上下文","Agnes AI 开源 33B 模型 Agnes-3.0-Flash Preview(Apache 2.0)。架构 3:1 混编,72 层仅 18 层 KV 缓存随上下文增长,单张 H200 跑 262,144 token 上下文。模型卡注明开源权重与 API 版(1M 上下文)不同源,跑分不混用。","Agnes AI 在 HuggingFace 上线了 33B 参数的 Agnes-3.0-Flash Preview,Apache 2.0 许可,权重可直接下载。这不是又一个\"Dense 模型换个名字\"的发布——它的注意力结构本身就是这篇模型卡最值得读的部分。\n\n## KV 缓存只长在 18 层\n\nAgnes-3.0-Flash Preview 是一个混合注意力解码器:每四层里,三层跑门控 delta 规则(gated delta rule)递归层,第四层才跑标准全局注意力,3:1 交替。全模型 72 层,54 层是递归层,18 层是全局注意力层。\n\n这个结构直接改变了推理时的内存曲线:delta 规则层的递归状态每层固定、与序列长度无关,只有那 18 层注意力层的 KV 缓存会随上下文增长——长上下文场景下用计算换内存的经典路线。模型卡规格全公开:递归状态 fp32 存储、hidden size 5120、词表 248,320、3 轴旋转位置编码。\n\n上下文窗口 262,144 token。硬件门槛:一张 H200 141GB 或 H100 80GB,bf16 权重约 66GB,建议主机内存 128GB 以上。一台单卡工作站就能跑满 262k 上下文的多模态模型,不需要集群。\n\n## 跑分:厂商自报口径\n\n模型卡给出对照表,但明确标注数字\"来自不同来源、不同 harness、不同模型快照,不构成受控对比\"。Agnes-3.0-Flash Preview 参考值:IFBench 74.20、GPQA Diamond 85.05、SciCode 38.08。对照列里,Qwen3.8-27B 的 GPQA 是 90.5、MiniMax M3 是 92.9——GPQA 上明显落后头部;IFBench 74.20 则高于 Qwen3.6-35B-A3B 的 64.4。\n\n换句话说,它的定位不是刷榜旗舰,而是\"单卡可跑的长上下文多模态基座\":支持文本、图像、视频理解,工具调用,chat template 带 high\u002Fmedium\u002Flow 三档推理力度。加载需要 trust_remote_code=True,transformers 5.12 起支持。\n\n## 一个罕见的诚实标注\n\n模型卡专门设了\"Model version clarification\"一节:这批开源权重是 Preview checkpoint(33B、262k 上下文),而 Artificial Analysis 榜单上的 Agnes 3.0 Flash 是另一个 production\u002FAPI checkpoint,1M token 上下文,\"其 benchmark 结果不应归到此处发布的 Preview 权重上\"。仓库最初就叫 Agnes-3.0-Flash,后来才显式加上 Preview 后缀澄清两者区别。\n\n开源版与 API 版同名不同能力,过去一年在这个行业里通常是用户自己踩坑才发现的;把区别写进模型卡第一节,并主动改仓库名,这个动作本身值得记一笔。\n\n## 所以呢\n\n截至发稿,HF 上已有 7 个社区量化版本(llama.cpp\u002FOllama 等生态),上月下载量 474——不算热,但\"架构混编 + Apache 2.0 + 单卡门槛\"这个组合,对做本地部署和长上下文应用的团队是个低成本试验对象。delta 规则与注意力 3:1 混编能否在更多尺度上复现,值得持续观察。\n\n引用:模型卡 https:\u002F\u002Fhuggingface.co\u002FAgnes-AI\u002FAgnes-3.0-Flash ;The Open Weights 报道 https:\u002F\u002Fwww.theopenweights.com\u002Fnews\u002Fagnes-3-0-flash-z86d","https:\u002F\u002Fhuggingface.co\u002FAgnes-AI\u002FAgnes-3.0-Flash","24d5c6c5-6573-4180-a1fd-f1459842d1af",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",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},"ac1d38a8-2724-4b5f-ac88-d41cf7ed0163","en","Agnes 3.0 Flash open weights: 33B, KV cache in 18 of 72 layers","Agnes AI open-sourced a 33B preview model (Apache 2.0) mixing delta-rule and attention layers 3:1; only 18 of 72 layers grow KV cache. Single H200 runs the 262k-token context.","Agnes AI has published Agnes-3.0-Flash Preview on HuggingFace: a 33B-parameter open-weights checkpoint under Apache 2.0. The most interesting part of this release is not the size but the attention structure itself.\n\n## KV cache grows in only 18 layers\n\nAgnes-3.0-Flash Preview is a hybrid-attention decoder. For every four layers, three run a gated delta rule (recurrent, with per-layer state independent of sequence length) and the fourth runs standard global attention, alternating 3:1. Of the 72 decoder layers, 54 are delta-rule recurrent layers and 18 are global attention layers.\n\nThis directly changes the memory curve at inference time: the recurrent state is fixed per layer regardless of sequence length, so only the 18 attention layers hold a KV cache that grows with context. The model card lists every spec: recurrent state in fp32, a causal conv (kernel 4) in front of each delta-rule layer, 16 key heads \u002F 48 value heads, hidden size 5,120, vocabulary of 248,320, and 3-axis rotary (text \u002F height \u002F width) with interleaved mrope sections 11:11:10.\n\nThe context window is 262,144 tokens. Recommended hardware from the model card: a single NVIDIA H200 141GB or H100 80GB, roughly 66GB for the bf16 checkpoint, 128GB+ host memory. A single-GPU workstation can run the full 262k-context multimodal model — no cluster required.\n\n## Benchmarks, vendor-reported\n\nThe model card includes a comparison table but explicitly notes the figures were \"compiled from different sources, harnesses, and model snapshots and do not constitute a controlled head-to-head comparison.\" Reference values for Agnes-3.0-Flash Preview: IFBench 74.20, GPQA Diamond 85.05, SciCode 38.08. In the same table, Qwen3.8-27B posts GPQA 90.5 and MiniMax M3 92.9 — Agnes clearly trails the leaders on GPQA; on IFBench, 74.20 beats Qwen3.6-35B-A3B's 64.4 and Kimi K2.5's 43.7.\n\nIn other words, this is not a leaderboard flagship but a single-GPU long-context multimodal base. It supports text, image and video understanding, tool calling via \u003Ctool_call>, and three reasoning-effort levels (high\u002Fmedium\u002Flow) in the chat template. Loading requires trust_remote_code=True, with transformers 5.12 or newer.\n\n## A rare honest label\n\nThe model card carries a dedicated \"Model version clarification\" section: these open weights are the Preview checkpoint (33B, 262k context), while the Agnes 3.0 Flash listed on Artificial Analysis is a different production\u002FAPI checkpoint with a 1M-token context — \"its benchmark results should not be attributed to the Preview weights released here.\" The repository was initially published without the Preview suffix and was renamed to make the distinction explicit.\n\nOpen-weight and API versions sharing a name with different capabilities is a trap users usually discover the hard way; writing the difference into the first section of the model card and renaming the repo deserves a note.\n\n## So what\n\nAs of this writing, 7 community quantizations already exist on HF (llama.cpp \u002F LM Studio \u002F Ollama \u002F Jan ecosystems), with 474 downloads last month — not hot, but the combination of hybrid architecture + Apache 2.0 + single-GPU entry makes it a cheap experiment target for local-deployment and long-context teams. Whether the 3:1 delta-rule\u002Fattention mix holds at other scales is worth watching.\n\nSources: model card https:\u002F\u002Fhuggingface.co\u002FAgnes-AI\u002FAgnes-3.0-Flash ; The Open Weights https:\u002F\u002Fwww.theopenweights.com\u002Fnews\u002Fagnes-3-0-flash-z86d","agnes-3-0-flash-preview-open-weights","2026-09-13T15:20:00Z","2026-09-13T15:08:06.438965Z","2026-09-13T15:08:06.438980Z",true,"agent",89,[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},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","minimax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"4c4a2a9e-f69b-4985-bd42-97ab2ef4e2ac","Spark-X2.5-4B 开源:4B 跑 1M 上下文,22 项基准打 9B 级 Qwen3.5","spark-x2-5-4b-apache-open-source","2026-09-16T01:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"51c13e24-8072-404c-a8d4-75c40cff05ee","Ling-3.0-flash-VL 开源：124B MoE 只激活 5.5B，视觉塞进 Agent 闭环","ling-3-0-flash-vl-open-weights","2026-09-15T13:18:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"108af093-b226-4372-9cf0-77323ffc5456","小鹏 X-AuT 给语音大模型剪枝:音频塔砍 4 层,车载推理提速 21.4%","xpeng-x-aut-audio-encoder-pruning","2026-09-12T19:06:47+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"365b770a-2cca-40a0-beb0-1eff823702c0","IBM 开源 Granite Time Series PatchTST-FM-r2:零样本 SOTA,Apache 2.0 商用许可","ibm-granite-patchtst-fm-r2-zero-shot-apache","2026-09-12T11:00:00+00:00"]