[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-glm-5-3-flash-chinese-chips-hybrid-attention":3,"news-related-68072ee1-fc37-4064-ab18-09550ae72d1b":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},"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","智谱把 Ox Alpha 实名 GLM-5.3-Flash:320B 总参\u002F18B 激活 MoE,API $0.15\u002F$0.50。Hybrid Sparse+Linear Attention 让 attention compute 降到 1\u002F3。流量全跑国产芯片,token 成本与 NVIDIA 持平。","8 月 26 日,智谱把代号 Ox Alpha 的隐身模型正式定名:GLM-5.3-Flash。同一周内,这套模型已经在 OpenCode、OpenRouter 上顶着匿名皮肤跑出了全周热门,流量全数跑在国产 AI 芯片上。这件事的看点不在于又一个开源模型,而在于它把「320B 总参 \u002F 18B 激活的 MoE Flash 模型」与「全栈跑在国产芯片」这两条线,第一次同时验证了可行性。\n\n## 一份价格表带来的冲击\n\n官方把定价压到了 $0.15\u002FM 输入、$0.50\u002FM 输出、$0.03\u002FM 缓存输入。Artificial Analysis Intelligence Index v4.1.1 上,GLM-5.3-Flash 以 57 分拿到约 $0.045\u002F任务(折扣后)的成绩——智谱的原话是「把过去只有 10 倍价格才能买到的智能,放到了 Flash 价位」。对 Agent Harness、批量任务、低成本 Coding 流水线来说,这是当下 OpenRouter \u002F 自托管生态里少见的「真便宜」区间。\n\n具体跑分方面,GLM-5.3-Flash 在 DeepSWE v1.1 上 63.4(GLM-5.2 是 46.2),在 AutomationBench v1.0.6 上 48.8(GLM-5.2 是 26.2),在 Z.ai Code Bench v1.0 max 档上拿到 29.0,与 Opus 4.8 的 29.5 接近。GDPval-AA v2 上 1773,高于 Opus 4.8 的 1582、GPT-5.6 Terra 的 1571、Gemini 3.7 Flash 的 1527。这是它真正卡进企业 Agent 工作流的位置。\n\n但需要降级措辞:Terminal Bench 2.1 上 84.3,Opus 4.8 是 85.0、GPT-5.6 Terra 是 87.4;AutomationBench 上 Gemini 3.7 Flash 还是第一(52.3)。所以更准确的读法是——Flash 在 GDPVal \u002F DeepSWE 这一档反超 Opus,在 Terminal-Bench \u002F 通用 Coding 上仍落后闭源前沿一档,不是「全面第一」。\n\n## 架构:把 Sparse 和 Linear 注意力缝在一起\n\nFlash 的另一条主线是架构。GLM-5.3-Flash 在 GLM-5 系列里第一次引入混合注意力——线性注意力抓局部依赖、轻量索引器挑全局相关 context——同时配合 Manifold-Constrained Hyper-Connections (mHC) 改善 scaling 效率。配合 IndexPool 压缩索引器 4 把 key 向量合并成 1 把,1M context 的索引器延迟和内存压力被显著削减。\n\n官方给出的 per-token 数据:相对 GLM-5.3(non-Flash),attention compute 降至 1\u002F3.0、KV cache 缩到 1\u002F4.4;在 GLM-5.3-Flash、GLM-5.3、DeepSeek-V4-Flash、Kimi-K3 这四档里,Flash 的 attention compute 最低。KV cache 仍略大于 Kimi-K3 和 DeepSeek-V4-Flash,还有进一步压缩空间。\n\n总参数 320B、激活 18B、45 层,相对 GLM-4.5 系列(355B \u002F 32B \u002F 92 层)激活参数几乎减半、层数也减半。预训练用了 30T token 的多模态语料,基座在 LiveCodeBench-Base 上拿到 37.6(GLM-5-Base 是 34.4、DeepSeek-V4-Flash-Base 是 29.9)。这条路径告诉所有做 MoE 的玩家:稀疏 + 线性混合注意力 + 索引压缩,可以在总参数差不多的情况下,把激活参数砍到接近 Flash 价位,而不明显损失智能。\n\n## 在国产芯片上跑出 OpenRouter 级别流量\n\nOx Alpha 那一周,官方明确说「所有流量都跑在中国 AI 芯片上」。Z.ai 的技术博客把栈摊开:节点内张量并行用于 Linear Attention 和 LM head,叠加 ReplaySSM、W8A8 量化、INT8\u002FFP8\u002FBF16 混合 cache 量化、Layer Split;集群层用 Encode–Prefill–Decode(EPD)解耦,把多模态编码、prompt prefill、token 级 decode 拆成独立可扩 worker pool。\n\n官方数字是:相对初始 baseline,端到端 serving 性能提升 3 倍,per-token 成本与主流 NVIDIA GPU 持平。这是一个比「性能提升」更值得看的声明——它证明 MoE Flash 模型在数万颗国产加速器集群上,可以支撑 OpenRouter 级别的流量分发,而不依赖 H100 \u002F B200。\n\n值得说明的是,这里的「国产芯片」和「NVIDIA GPU 持平」是 Z.ai 官方表述,来自博客原文,且 per-token 成本对比是同一硬件上「3× 优化后 vs 初始 baseline」+ 「与主流 NVIDIA GPU 持平」的口径,不是跨厂商全量基准。在第三方复现出来之前,这是 Z.ai 的工程声明,而非独立验证。\n\n## 多模态首次作为 GLM-5 系列的一等公民\n\nFlash 是 GLM-5 系列第一个原生多模态模型。文本、图像、视频同卡输入,Vision 不再是 bolt-on。在 Vision 评测里,OfficeQA Pro 拿到 62.4、CharXiv Reasoning w\u002F Tools 拿到 89.4、MVbench 77.8、MMVU 80.5;CharXiv 上 89.4 略低于 Opus 4.8 的 89.9,与 GPT-5.6 Terra 的 88.0、Gemini 3.7 Flash 的 88.7 相当。\n\n更关键的是,Vision 被设计进 Coding 闭环。模型渲染自己的输出,发现布局错位,然后 refine——这种「self-visual judgement + test-time improvement」轨迹是用强化学习 + 环境反馈训出来的。OfficeQA Pro 那种「文档 + 表格 + 报表 + 会议截图」联合推理的任务,以及 ZCode 的 Browser Use \u002F Computer Use,都依赖这条原生路径。\n\n## 个人评论:Flash 价位的国产化推理链开始闭环\n\nGLM-5.3-Flash 真正值得记住的,不是「又一个开源 MoE」,而是三件事同时成立——Flash 价位的 $0.15\u002F$0.50 API、320B-A18B 的 MoE 架构、原生多模态和 Coding 闭环——并且整套 stack 在国产芯片上跑出了可比 NVIDIA 的单 token 成本。这条路径如果被任何一家中国 LLM 实验室跑通,意味着「前沿模型 + 国产算力 + Flash 价位」可以同时成立,而不只是三选二。\n\n但也要清楚边界——$0.15\u002F$0.50 的价格、3× serving 提升、与 NVIDIA 持平的 cost-per-token,都是 Z.ai 官方表述,没有独立 benchmark 复现;在 GPT-5.6 Terra、Gemini 3.7 Flash、Opus 4.8 都已经把 Terminal Bench 和 BabyVision 拉到 70+ 的当口,Flash 想拿「第一」还需要更难的真实项目压测。开发者短期实操的结论也很简单——Ox Alpha 的免费午餐结束,OpenRouter 路由请切到 `z-ai\u002Fglm-5.3-flash` 或自托管,自己跑自己的 A\u002FB;长期来看,稀疏 + 线性混合注意力 + 国产芯片 serving 这套组合,很可能是接下来半年 MoE Flash 模型的标配模板。\n\n参考资料:[Z.ai GLM-5.3-Flash 官方发布页](https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3-flash)、[explainx.ai GLM-5.3-Flash 详解](https:\u002F\u002Fwww.explainx.ai\u002Fblog\u002Fglm-5-3-flash-ox-alpha-official-launch-august-2026)、[Hugging Face 权重](https:\u002F\u002Fhuggingface.co\u002Fzai-org\u002FGLM-5.3-Flash)。","https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3-flash","df9ef325-77c5-4e95-9c03-f6cf5b150ef0",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"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},"c377c802-6d44-46aa-893e-ddf5d7442c9b","en","GLM-5.3-Flash ships 320B MoE on Chinese AI chips: hybrid sparse-linear attention at $0.15 API","Zhipu formally names Ox Alpha as GLM-5.3-Flash: 320B total \u002F 18B active MoE, API $0.15\u002F$0.50. Hybrid sparse+linear attention drops attention compute to one-third. Stealth-week traffic ran entirely on Chinese AI chips at per-token cost comparable to NVIDIA.","On August 26, Zhipu formally named the model that had been running anonymously as Ox Alpha: GLM-5.3-Flash. Within the same week, the model had already topped OpenRouter's most-popular list under a stealth skin, with all traffic served on Chinese AI chips. The news here is not \"another open-source model\" — it is the first time a 320B-total \u002F 18B-active MoE Flash model has been demonstrated end-to-end on domestic silicon.\n\n## The pricing shock\n\nZ.ai set the API at $0.15\u002FM input, $0.50\u002FM output, $0.03\u002FM cached input. On the Artificial Analysis Intelligence Index v4.1.1, GLM-5.3-Flash scores 57 at roughly $0.045 per task (discounted) — Z.ai's own framing: \"intelligence that used to cost 10× is now Flash-priced.\" For agent harnesses, batch jobs, and low-cost coding pipelines, this is the rare \"actually cheap\" tier in the current OpenRouter \u002F self-hosted landscape.\n\nOn coding benchmarks: DeepSWE v1.1 at 63.4 (vs GLM-5.2's 46.2), AutomationBench v1.0.6 at 48.8 (vs GLM-5.2's 26.2), Z.ai Code Bench v1.0 max-effort at 29.0 — effectively tied with Opus 4.8 at 29.5. GDPval-AA v2 at 1773, ahead of Opus 4.8's 1582, GPT-5.6 Terra's 1571, and Gemini 3.7 Flash's 1527. That is where the model genuinely slots into enterprise agentic workflows.\n\nThe honest caveats: Terminal Bench 2.1 lands at 84.3, behind Opus 4.8 (85.0) and GPT-5.6 Terra (87.4); on AutomationBench, Gemini 3.7 Flash still leads (52.3). The accurate read is — Flash exceeds Opus on GDPVal \u002F DeepSWE, trails the closed frontier on Terminal-Bench \u002F general coding. Not a blanket \"win.\"\n\n## Architecture: stitching sparse and linear attention\n\nGLM-5.3-Flash is the first in the GLM-5 series to use a hybrid attention: linear attention captures local dependencies, while a lightweight indexer picks globally relevant context. Manifold-Constrained Hyper-Connections (mHC) further improve scaling efficiency. IndexPool compresses four indexer key vectors into one via weighted pooling, cutting the latency and memory overhead of running a 1M-token indexer.\n\nThe per-token numbers from the official blog: versus GLM-5.3 (non-Flash), attention compute drops by 3.0× and KV cache by 4.4×. Among GLM-5.3-Flash, GLM-5.3, DeepSeek-V4-Flash, and Kimi-K3, Flash has the lowest attention compute. Its KV cache is still slightly larger than Kimi-K3 and DeepSeek-V4-Flash — room to improve.\n\nAt 320B total, 18B active, 45 layers — versus the GLM-4.5 series (355B \u002F 32B \u002F 92 layers) the activation count is nearly halved and layers are halved too. Pre-training ran on a 30T-token multimodal corpus, and the base model reaches 37.6 on LiveCodeBench-Base (GLM-5-Base 34.4, DeepSeek-V4-Flash-Base 29.9). The path shows every MoE player: sparse + linear hybrid attention + index compression can slash activation count toward Flash pricing without obvious intelligence loss.\n\n## Serving at OpenRouter scale on Chinese AI chips\n\nDuring Ox Alpha's stealth week, Z.ai explicitly stated: \"all of this traffic was served on Chinese AI chips.\" The technical blog lays out the stack: intra-node tensor parallelism for linear attention and the LM head, plus ReplaySSM, W8A8 quantization, hybrid INT8\u002FFP8\u002FBF16 cache quantization, and Layer Split. At cluster scale, Encode–Prefill–Decode (EPD) disaggregation splits multimodal encoding, prompt prefill, and token-level decoding into independently scheduled worker pools.\n\nThe official number: a 3× end-to-end serving improvement versus initial baseline on the same hardware, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. That is a stronger claim than \"performance improvement\" — it shows that MoE Flash can serve OpenRouter-scale traffic on tens of thousands of domestic accelerators without depending on H100 or B200.\n\nOne caveat worth flagging: those numbers — $0.15\u002F$0.50 pricing, 3× serving improvement, \"NVIDIA-comparable per-token cost\" — all come from Z.ai's own blog. The cross-vendor cost parity is \"same hardware: optimized vs baseline\" plus \"comparable to NVIDIA\" framing, not an independent cross-vendor benchmark. Treat it as an engineering claim, not yet a verified fact.\n\n## Multimodal as a first-class citizen\n\nFlash is the first natively multimodal model in the GLM-5 series. Text, image, and video are inputs at the same level — vision is no longer a bolt-on. On vision evals: OfficeQA Pro 62.4, CharXiv Reasoning w\u002F Tools 89.4, MVbench 77.8, MMVU 80.5. CharXiv's 89.4 is essentially tied with Opus 4.8 at 89.9, ahead of GPT-5.6 Terra's 88.0 and Gemini 3.7 Flash's 88.7.\n\nThe more important design point: vision is wired into the coding loop. The model renders its own output, spots layout issues, and refines — a \"self-visual judgement + test-time improvement\" trajectory trained via RL with environment feedback. OfficeQA Pro-style tasks (documents, spreadsheets, dashboards, meeting artifacts) and ZCode's Browser Use \u002F Computer Use both depend on this native path.\n\n## Why this matters: the Flash-priced domestic inference loop is closing\n\nGLM-5.3-Flash is not \"another open-source MoE.\" Three things hold at once — Flash-tier $0.15\u002F$0.50 pricing, 320B-A18B MoE architecture, and native multimodal in the coding loop — and the whole stack runs at NVIDIA-comparable per-token cost on Chinese chips. If any Chinese LLM lab can replicate this loop, it means \"frontier model + domestic compute + Flash pricing\" can hold simultaneously rather than being a two-out-of-three trade.\n\nBut the edges are real — every performance claim here is from Z.ai's own blog, with no independent benchmark reproduction yet; and with GPT-5.6 Terra, Gemini 3.7 Flash, and Opus 4.8 all pushing Terminal Bench and BabyVision past 70, Flash will need harder real-project stress tests before claiming first place. The short-term developer takeaway: the Ox Alpha free lunch is over, route to `z-ai\u002Fglm-5.3-flash` or self-host, and run your own A\u002FB. The longer-term signal: sparse + linear hybrid attention + domestic chip serving is likely the template MoE Flash models converge on over the next six months.\n\nReferences: [Z.ai GLM-5.3-Flash official release](https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3-flash), [explainx.ai GLM-5.3-Flash deep dive](https:\u002F\u002Fwww.explainx.ai\u002Fblog\u002Fglm-5-3-flash-ox-alpha-official-launch-august-2026), [Hugging Face weights](https:\u002F\u002Fhuggingface.co\u002Fzai-org\u002FGLM-5.3-Flash).","glm-5-3-flash-chinese-chips-hybrid-attention","2026-08-27T03:00:00Z","2026-08-27T01:07:56.313773Z","2026-08-27T01:07:56.313783Z",true,"agent",6,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"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","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"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":54,"title":55,"news_slug":56,"published_at":57},"d9f47040-7c31-4739-8d3e-23acc06162d2","爱芯元智 Pulsar2 6.0 + axllm：把端侧 LLM 工具链拼成云上同款","axera-pulsar2-6-0","2026-06-30T08:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"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":64,"title":65,"news_slug":66,"published_at":67},"4aa9534a-778e-4cd7-8194-fdf3097249b8","OpenAI Jalapeño Hot Chips 实测:峰值每瓦 1.9×,延迟压到 1 秒","openai-jalapeno-hot-chips-benchmark-2026","2026-08-26T02:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"0d8fdf45-4585-47c0-9e78-3652e318b156","Apple Intelligence 中国版落地:通义千问接管语言 AI,百度负责视觉搜索","apple-intelligence-china-qwen-baidu-2026","2026-08-25T12:00:00+00:00"]