[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepseek-v4-ga-time-of-use-pricing":3,"news-related-894787b7-8799-4c22-8e20-5e04db8f7181":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"894787b7-8799-4c22-8e20-5e04db8f7181","DeepSeek V4 GA 把分时电价塞进大模型 API:CSA + HCA 撑起 1M 上下文,峰值翻倍只是商业答案","7 月 20 日 DeepSeek V4 正式 GA,沿用预览期 1.6T MoE 架构,核心新增是面向工业级推理的\"峰谷定价\"——北京工作日 09:00-12:00 与 14:00-18:00 输出一倍,off-peak 仅 $0.87\u002FM。配合 CSA + HCA 把 KV 缓存压到 V3.2 的 10%,V4-Pro 终于让 1M 上下文从 spec sheet 走向可规模化部署。\n\nDeepSeek V4 在 4 月以 MIT 协议开源时,大家关注的是 1.6T 总参 \u002F 49B 激活的 MoE 架构和 1M 上下文;但真正卡生产部署的不是参数,是那一行 spec sheet 上写着\"1M 上下文\"、实际却烧光显存撑不开。\n\n7 月 20 日 V4 GA 把架构故事和商业故事拆开回答了两个不同的问题:\n\n第一,1M 上下文凭什么便宜?答案在注意力层的工程重组:用 CSA(Compressed Sparse Attention,4× softmax-gated pooling + FP4 闪电索引器 top-1024 sparse)与 HCA(Heavily Compressed Attention,128× 压缩后做全局 dense attention)按 transformer 层交替铺开。1M tokens 时 FLOPs 仅 V3.2 的 27%,KV cache 缩到 10%(Flash 版本 7%)。这意味着 V4-Pro 的 384K max-output + 1M 上下文,是第一次真正可以按 token 经济地定价的 1M 上下文。\n\n第二,API 凭什么分时?答案是商业模式升级:GA 引入峰谷输电价——off-peak 仅 $0.87\u002FM,peak 双倍至 $1.74\u002FM。即便翻倍,仍比 Claude Opus 4.8($15\u002FM)和 GPT-5.6 Sol(约 $15\u002FM)便宜一个数量级,同时按\"Strategy & Operations\"基准折算成本,V4-Pro 单任务仅 $0.03,Claude Fable 5 是 $3.48,116× 成本差距。\n\nSWE-bench Verified 上 V4-Pro 拿到 80.6%,刷新开源权重纪录并打平 Gemini 3.1 Pro;LiveCodeBench 93.5%、Codeforces Elo 3206 同样领跑。把这两个数字放进 API 经济账里:开源模型第一次证明,在工程成本维度,闭源 frontier 不再是默认选项。\n\n工程上的\"对齐+加速\"在同一表征空间实现,大概率会成为下一代开源 LLM 的标配范式——而 DeepSeek 已经把这条路径走到了 GA。","https:\u002F\u002Fmacgpu.com\u002Fen\u002Fblog\u002F2026-0720-deepseek-v4-full-release-pricing-benchmarks.html","22403a2a-1cbd-49c2-b55b-e51164f5f4c2",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"aabbac8a-b4db-475c-9222-7df485361d57","en","DeepSeek V4 GA folds time-of-use pricing into the API","DeepSeek V4 went GA on July 20, sticking with the preview's 1.6T MoE architecture, with the core addition being \"peak\u002Fvalley pricing\" for industrial-grade inference — Beijing working hours 09:00–12:00 and 14:00–18:00 charge 1x for output, off-peak only $0.87\u002FM. Combined with CSA + HCA pushing KV cache down to 10% of V3.2, V4-Pro finally makes 1M context move from spec sheet to scalable deployment. When DeepSeek V4 open-sourced under MIT in April, the focus was on the 1.6T total \u002F 49B active MoE architecture and 1M context; but the real blocker for production deployment wasn't the parameters — it was the spec-sheet \"1M context\" that in practice burned out VRAM. V4 GA on July 20 split the architecture story and the business story to answer two different questions: First, why is 1M context cheap? The answer is in the engineering restructuring of the attention layers: CSA (Compressed Sparse Attention, 4x softmax-gated pooling + FP4 lightning indexer top-1024 sparse) and HCA (Heavily Compressed Attention, 128x compression then global dense attention) alternate by transformer layer. At 1M tokens, FLOPs are only 27% of V3.2, and KV cache shrinks to 10% (Flash version 7%). This means V4-Pro's 384K max output + 1M context is the first 1M context that can actually be priced per-token economically. Second, why is the API time-of-use? It's a business-model upgrade: GA introduces peak\u002Fvalley output pricing — off-peak only $0.87\u002FM, peak doubles to $1.74\u002FM. Even doubled, it's still an order of magnitude cheaper than Claude Opus 4.8 ($15\u002FM) and GPT-5.6 Sol (~$15\u002FM), and by the \"Strategy & Operations\" benchmark the per-task cost is just $0.03, while Claude Fable 5 is $3.48 — a 116x cost gap. On SWE-bench Verified, V4-Pro scores 80.6%, refreshing the open-weight record and tying Gemini 3.1 Pro; LiveCodeBench 93.5%, Codeforces Elo 3206 also lead. Put those two numbers into the API economics ledger: for the first time, open-source models prove that on engineering cost, closed-source frontier is no longer the default option. Engineering-side \"alignment + acceleration\" in the same representation space will almost certainly become the standard paradigm for the next generation of open-source LLMs — and DeepSeek has already taken this path all the way to GA.","deepseek-v4-ga-time-of-use-pricing","2026-07-24T02:00:00Z","2026-07-24T04:06:38.056918Z","2026-08-19T02:08:40.142862Z",true,"agent",91,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"cac485ab-a429-4ebc-88c6-a1f924f978ff","AWS 开源 KeysAndValues:微调时就让模型学会“遗忘”,单张 A100 撑住 128K","aws-keysvalues-sparse-attention-finetuning","2026-08-26T05:20:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"c94bdf86-5de9-49fe-8c98-0f5c47611bfe","SGLang v0.5.18 发布:大模型冷启动提速 2.38 倍,710 个 PR 都改了什么","sglang-v0-5-18-cold-start-2-38x","2026-08-24T23:15:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"92433e6b-113a-4ada-af77-fbb8995a9850","LFM2.5-DSpark 开源:300M 草稿模型让端侧推理快 2.87 倍,输出零损耗","lfm2-5-dspark-draft-models","2026-08-21T21:10:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"926f89fc-5ed5-4170-bfac-931d3a31b6a4","腾讯混元开源 AngelSpec 投机解码框架：DFly 在 Hy3-A21B 上取得 1.98–2.40× 加速","tencent-angelspec-spec-decoding-hy3-dfly","2026-07-30T00:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"518d0be2-1220-4920-98fd-dee9df27a43d","VIDRAFT VKUE 把同一份 34.7B 稀疏 MoE 权重从 B200 跑到裸 CPU","vidraft-vkue-cpu-moe","2026-07-12T10:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":62},"f8a33525-3846-46e2-9f88-05be407dbe2f","Liquid AI 用 Antidoom 把推理模型的 doom loop 压到 1%：一次训练只改 1 个 token","liquid-ai-antidoom"]