[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-glm-5-3-flash-hybrid-attention-architecture":3,"news-related-804ab59a-a8d6-4b61-bf74-8f6f2bdae83c":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"804ab59a-a8d6-4b61-bf74-8f6f2bdae83c","智谱把 Flash 做成一件正经事:一次说清 GLM-5.3-Flash 的架构和 benchmark 真相","智谱开源 GLM-5.3-Flash:320B\u002F18B MoE、1M 上下文、MIT 协议。混合 KDA+NoPE 稀疏 MLA 把 attention 算力压到 1\u002F3。Z.ai Code Bench 距 Opus 4.8 仅 0.5 分,AA 给到 57,视觉是弱项。","## 一次说清:这台机器的注意力架构和 benchmark 真相到底长什么样\n\n8 月 26 日,智谱(Z.ai)把上周顶着「Ox Alpha」匿名皮肤跑了一周的 GLM-5.3-Flash 开源上架。同一时间 SiliconANGLE、MarkTechPost 都做了同步报道。SiliconANGLE 的重点是「开源权重」,MarkTechPost 的重点是「Flash 价位和混合注意力栈」,今天这篇想做一件更难一点的事:把这台机器的注意力架构拆开,看 Z.ai 给的 benchmark 数字到底意味着什么,以及「接近 Opus 4.8」这种话在独立评分体系下还站不站得住。\n\n## 架构:GLM 系列第一次用 KDA + NoPE 稀疏 MLA 混搭\n\nGLM-5.3-Flash 不是一个 320B\u002F18B 的常规 MoE,它把注意力机制拆成两段做了第一次「混搭」,这套方案在 MarkTechPost 的拆解里有非常清晰的描述。45 层语言模型,**KDA 线性注意力层**负责局部依赖,**NoPE 稀疏 MLA 层**负责检索全局相关 context。每个 token 路由通过 **8 个专家,总专家数 288**。原生 FP8 权重外加一个 MTP draft 层。\n\n这套架构真正解决的问题是 1M 上下文下的检索瓶颈。智谱为此专门做了一个叫 **IndexPool** 的机制——把 indexer 的 key 向量分组做加权池化,把检索成本压下去。最终官方数字是「相对 GLM-5.3,attention 计算量减少约 3 倍,KV cache 缩小 4.4 倍」。第三个改动是 **mHC(Manifold-Constrained Hyper-Connections)**——约束梯度流过的拓扑,MarkTechPost 引用的官方说法是「相对 GLM-4.5,在相近总参数规模下,激活参数和层数都大约减半」。\n\n直白讲:同样 1M context,显存占用从「几乎装不下」变成「一台 8 卡 Hopper 节点能跑」——这正是 vLLM recipe 上写「约 306 GiB FP8 权重,Hopper 及以上」的现实依据([vLLM recipe](https:\u002F\u002Frecipes.vllm.ai\u002Fzai-org\u002FGLM-5.3-Flash))。\n\n## Benchmark:官方数字漂亮,「接近 Opus 4.8」要看哪个分数\n\n先把 Z.ai 自己给的核心数字摆出来,全部来自 8 月 14 日发布的官方博客([z.ai\u002Fblog\u002Fglm-5.3](https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3))。\n\n- **Terminal-Bench 3.0**:GLM-5.3 = 28.3,Opus 4.8 = 21.1,Fable 5 = 33.7,GPT-5.6 Sol = 34.6\n- **DeepSWE v1.1**:GLM-5.3 = 66.9,Opus 4.8 = 58.0,Fable 5 = 69.7,GPT-5.6 Sol = 72.7\n- **Z.ai Code Bench max effort**:GLM-5.3 = 34.5%(约 75K 输出 token),Opus 4.8 落后,Fable 5 = 39.5% 仍第一\n- **CyberGym**:GLM-5.3 = 84.5%,Fable 5 = 83.8%,GPT-5.6 Sol = 83.6%\n- **ExploitBench**:GLM-5.3 = 54.4%,Fable 5 = 78.0%,GPT-5.6 Sol = 76.5%\n\nFlash 这一档(Z.ai Code Bench v1.0 max):**29.0 对 Opus 4.8 的 29.5**,只差 0.5 分。听起来真挺接近。\n\n但如果换成 **Artificial Analysis** 这个独立评分体系([artificialanalysis.ai\u002Fmodels\u002Fglm-5-3-flash](https:\u002F\u002Fartificialanalysis.ai\u002Fmodels\u002Fglm-5-3-flash)),GLM-5.3-Flash 在 Intelligence Index v4.1.1 上拿到 **57 分**,Z.ai API 上跑出 **48.7 输出 token\u002F秒** 和 **1.52 秒 TTFT**。57 分意味着它是「智能\u002F价格比很强,但绝对智能水平还是 Flash 档」—— 这跟 Opus 4.8 不在一个段位。\n\n## 真正的弱项:视觉\n\nFlash 在 SiliconANGLE 报道中被标榜为「native 多模态」,支持图像和视频输入,但 MarkTechPost 单独点出了它的视觉短板的:**在 BabyVision 和 MVbench 上落后 Gemini 3.7 Flash**。换句话说,这个新模型的「1M context + 多模态」在文本和代码侧是真的好用,视觉部分还没有到能跟 Gemini 3.7 Flash 抢地盘的阶段。\n\n## 价格与可触达性\n\n- **API**:**$0.15\u002FM input、$0.03\u002FM cached input、$0.50\u002FM output**(标准档),Z.ai 称在折扣档下能做到 **$0.045\u002Ftask**\n- **GLM Coding Plan 三档**(Lite $18 \u002F Pro $80 \u002F Max $168)全部铺开,**配额是 GLM-5.3 的 3 倍**\n- **本地推理**:SGLang、vLLM、TokenSpeed、KTransformers 全部支持\n\n## 所以呢\n\n这件事的真正信号,不是「又一个开源模型」。它的真正信号是:**国产 Flash 档模型第一次把 1M 上下文 + 原生多模态 + MIT 开源 + 混合注意力栈,同时压在 $0.15 的 API 价位上**。在「同等智能成本」的赛道上,智谱给出了另一条答案——不靠更小模型降本,而靠架构创新让 320B MoE 跑得起 Flash 价位。\n\n至于「接近 Opus 4.8」—— 在 Z.ai 自己设计的 Code Bench 上是接近,在 Artificial Analysis 这种独立综合评分上还有明显距离。看哪个分数,决定你看到的是「旗舰级」还是「Flash 级」。","https:\u002F\u002Fwww.marktechpost.com\u002F2026\u002F08\u002F26\u002Fz-ai-releases-glm-5-3-flash-a-320b-a18b-natively-multimodal-moe-with-a-1m-token-context\u002F","df9ef325-77c5-4e95-9c03-f6cf5b150ef0",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"e1b60f52-fa2c-41b9-9b1e-4c211391b1cb","en","What Z.ai's hybrid attention actually buys you at $0.15\u002FM","Z.ai open-sourced GLM-5.3-Flash on August 26: 320B\u002F18B MoE, 1M context, MIT license. The hybrid KDA linear + NoPE sparse MLA stack cuts attention compute by ~3x and KV cache by 4.4x. Z.ai Code Bench sits 0.5 points behind Opus 4.8; Artificial Analysis scores it 57 on the Intelligence Index. Vision trails Gemini 3.7 Flash.","## What the architecture actually buys you, and where the benchmarks disagree\n\nOn August 26, Z.ai (Zhipu) open-sourced GLM-5.3-Flash — the same model that ran under the \"Ox Alpha\" anonymous skin on OpenRouter and OpenCode for the previous week. Both SiliconANGLE and MarkTechPost covered the release simultaneously. SiliconANGLE focused on \"open weights\"; MarkTechPost focused on the \"Flash-tier price and hybrid attention stack.\" This piece tries to do something harder: open up the attention architecture, work through what Z.ai's benchmark numbers actually mean, and check whether \"approaching Opus 4.8\" still holds up under an independent scoring framework.\n\n## Architecture: GLM series uses KDA + NoPE sparse MLA hybrid for the first time\n\nGLM-5.3-Flash is not a routine 320B\u002F18B MoE. It splits the attention mechanism into two segments in a hybrid arrangement, clearly described in MarkTechPost's teardown. The 45-layer language model interleaves **KDA linear-attention layers** (handling local dependencies) with **NoPE sparse MLA layers** (retrieving globally relevant context). Each token is routed through **8 of 288 experts**. Native FP8 weights ship alongside one MTP draft layer.\n\nThe real problem this architecture solves is the retrieval bottleneck at 1M-token context. Z.ai built a mechanism called **IndexPool** specifically for this — it applies weighted pooling to groups of indexer key vectors to keep retrieval cost down. The official figures: \"compared with GLM-5.3, attention compute is reduced by approximately 3× and KV cache is 4.4× smaller.\" The third change is **mHC (Manifold-Constrained Hyper-Connections)** — constraining the topology gradients flow through. Per MarkTechPost citing official sources, \"compared with GLM-4.5, at similar total parameter counts, both activated parameters and layer count are roughly halved.\"\n\nPut plainly: at 1M context, memory usage shifts from \"barely fits\" to \"fits on a single 8-GPU Hopper node\" — which is exactly why the vLLM recipe page states \"approximately 306 GiB of FP8 weights, Hopper and above\" ([vLLM recipe](https:\u002F\u002Frecipes.vllm.ai\u002Fzai-org\u002FGLM-5.3-Flash)).\n\n## Benchmarks: official numbers look strong, but \"approaching Opus 4.8\" depends on which score\n\nFirst, the core numbers from Z.ai's own official blog post from August 14 ([z.ai\u002Fblog\u002Fglm-5.3](https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3)):\n\n- **Terminal-Bench 3.0**: GLM-5.3 = 28.3, Opus 4.8 = 21.1, Fable 5 = 33.7, GPT-5.6 Sol = 34.6\n- **DeepSWE v1.1**: GLM-5.3 = 66.9, Opus 4.8 = 58.0, Fable 5 = 69.7, GPT-5.6 Sol = 72.7\n- **Z.ai Code Bench (max effort)**: GLM-5.3 = 34.5% (~75K output tokens), Opus 4.8 trails, Fable 5 = 39.5% still first\n- **CyberGym**: GLM-5.3 = 84.5%, Fable 5 = 83.8%, GPT-5.6 Sol = 83.6%\n- **ExploitBench**: GLM-5.3 = 54.4%, Fable 5 = 78.0%, GPT-5.6 Sol = 76.5%\n\nFor the Flash tier (Z.ai Code Bench v1.0 max): **29.0 vs Opus 4.8's 29.5** — a gap of only 0.5 points. Sounds quite close.\n\nBut switch to the **Artificial Analysis** independent scoring framework ([artificialanalysis.ai\u002Fmodels\u002Fglm-5-3-flash](https:\u002F\u002Fartificialanalysis.ai\u002Fmodels\u002Fglm-5-3-flash)), and GLM-5.3-Flash scores **57 on the Intelligence Index v4.1.1**, with **48.7 output tokens\u002Fsecond** and **1.52s TTFT** on Z.ai's API. A score of 57 means it has strong intelligence-per-dollar, but its absolute intelligence level is still Flash-tier — not in the same league as Opus 4.8.\n\n## The real weakness: vision\n\nSiliconANGLE's coverage bills the model as \"natively multimodal,\" with image and video input supported. But MarkTechPost specifically called out its vision weak spot: **it trails Gemini 3.7 Flash on BabyVision and MVbench**. In other words, this new model's \"1M context + multimodal\" combo genuinely works well on text and code, but the vision side has not yet reached the level where it can compete with Gemini 3.7 Flash.\n\n## Price and accessibility\n\n- **API**: **$0.15\u002FM input, $0.03\u002FM cached input, $0.50\u002FM output** (standard tier); Z.ai reports **$0.045\u002Ftask** at the discounted tier\n- **GLM Coding Plan three tiers** (Lite $18 \u002F Pro $80 \u002F Max $168) all deployed, with **3× the quota of GLM-5.3**\n- **Local inference**: SGLang, vLLM, TokenSpeed, and KTransformers all supported\n\n## So what\n\nThe real signal here is not \"another open-source model.\" The signal is this: **a Chinese Flash-tier model is the first to combine 1M context + native multimodality + MIT open-source + hybrid attention stack, all at a $0.15 API price point.** On the \"equal intelligence for the same cost\" track, Zhipu offers a different answer — not by shrinking the model to cut costs, but by using architectural innovation to make a 320B MoE viable at Flash-tier pricing.\n\nAs for the \"near Opus 4.8\" claim — on Z.ai's own Code Bench it is close; on Artificial Analysis's independent composite scoring there is still a clear gap. Which score you look at decides whether you read it as \"flagship-tier\" or \"Flash-tier.\"","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00Z","2026-08-27T03:07:03.680566Z","2026-08-27T03:07:03.680573Z",true,"agent",9,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"f6e4aab0-7693-4c2c-bb66-c1641fc2cc3e","Ox Alpha 谜底揭晓:智谱 GLM-5.3-Flash,MIT 开源 320B MoE","ox-alpha-glm-5-3-flash-reveal","2026-08-27T13:30:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"68072ee1-fc37-4064-ab18-09550ae72d1b","GLM-5.3-Flash 把 320B MoE 跑在国产芯片上:Flash 价位和 $0.15 API 的混合注意力栈","glm-5-3-flash-chinese-chips-hybrid-attention","2026-08-27T03:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"7958a2f1-028c-4b4e-b134-0d5de9afc1c1","Motif 3 收官:韩国 314B MoE 改用 MIT 许可,从零起步架构首次面向商用","motif-3-mit-license-sovereign-ai","2026-08-24T00:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"491f4904-c854-4925-b3e3-e34b8afd5e50","KDA+MLA 混合栈下沉到 1.3B 激活:Ling-3.0-tiny 把 MoE 端侧化,INT4 跑出 115 tok\u002Fs","ling-3-tiny-kda-mla-edge-deployment","2026-08-18T00:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"b0183d10-bcfd-44ed-a178-a2c813f10b69","国家超算互联网AI社区上线Kimi K3:2.8万亿参数MoE一键调用,开源大模型有了国产算力底座","kimi-k3-cnsc-internet-launch","2026-07-28T09:30:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"3d8b9b1a-e038-466f-9b6b-304f911e35a7","Kimi K3 开源三件套 MoonEP\u002FFlashKDA\u002FAgentEnv:Moonshot 把 2.8T MoE 训练栈完整交底","kimi-k3-moonep-flashkda-agentenv","2026-07-28T04:30:00+00:00"]