[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepseek-v4-flash-0731-post-training-ga":3,"news-related-96f87320-5237-44cc-9919-dcee43a6ef80":38},{"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},"96f87320-5237-44cc-9919-dcee43a6ef80","DeepSeek-V4-Flash-0731 正式版 API 公测:只换后训练不换权重,开源 LLM 进入轻量迭代节奏","DeepSeek 发布 V4-Flash 正式版 API,与 preview 版权重完全一致,仅重新后训练。开源 LLM 在权重不动的前提下靠 RLHF\u002F指令微调做快速版本号滚动,V4-Pro 也即将发布。","## 只动后训练,不动权重:DeepSeek-V4-Flash 正式版 API 上线\n\n7 月 31 日,DeepSeek 在更新日志中宣布 **DeepSeek-V4-Flash-0731 正式版 API** 正式上线公测,同步预告了 V4-Pro 正式版即将发布。和 V4-Flash-preview 相比,正式版的 **模型结构、参数量、权重** 都保持不变,只重新做了一轮后训练(post-training)。也就是说,这次升级没有新 checkpoint、没有新架构,模型文件本身一字节都没改。\n\n这正是开源 LLM 进入工程化运维阶段后的典型打法:权重冻结 + 后训练迭代。\n\n### 为什么可以只换后训练\n\nV4-Flash 走的是 DeepSeek 标志性的 **MoE + 稀疏注意力 + 长上下文(1M tokens)** 路线。在 5 月 V4 GA 的技术披露中,DeepSeek 已经把 **CSA(Compressed Sparse Attention)+ HCA(Hierarchical Compressed Attention)** 两条主线拼到了同一套推理栈里,把 1M 上下文下的注意力计算量压到了传统密集注意力的 27% 左右,峰值吞吐比 V3 翻倍,但 API 价格长期按分时电价思路压得很低。\n\n后训练(post-training)成本远低于预训练:\n\n- **预训练**:数千卡 · 数十天 · 数百万美元\n- **后训练(RLHF \u002F SFT \u002F RL-on-CoT)**:几十到几百卡 · 数天到数周 · 成本量级低 2 到 3 个数量级\n\nFlash-preview 之所以单独命名,是因为它的权重是上一轮 RLHF\u002FSFT 流水线收尾时的副产物,DeepSeek 想拿真实流量再跑一轮对齐。**0731 这个版本号本身就是一次完整后训练迭代的收尾标记**,而不是任何架构升级。\n\n### 这种打法的三个收益\n\n1. **快速滚动版本号**:开源 LLM 不能再像过去那样每次都喊训练成本数百万美元。用户每天在生产环境跑推理,质量 regression 必须被快速修正。Preview 到正式版的无权重升级,可以在几天内通过 API 推给所有用户。\n\n2. **行为校准而不打扰基座**:V4-Flash 在多语言、长上下文、代码、推理上是开源同价位里最强的(参考 Developers Digest 三模型对比里的 DeepSeek V4 部分)。这些能力写在权重里,**改权重 = 承担 regression 风险**;改对齐层 = 改说话方式,基座推理能力保留。\n\n3. **成本曲线平滑**:preview 阶段用户量大,正式版意味着 DeepSeek 已经把对齐数据 \u002F reward model \u002F RL 训练流水线都跑稳,**进入边际成本极低的稳定运营态**。这是为什么 V4-Pro 正式版可以尽快发布——同一条后训练管线可以无缝切换到 Pro 权重上。\n\n### 留给行业的两个问题\n\n- **复现性问题**:只看 API 行为,用户不知道哪些是权重贡献、哪些是对齐贡献。一旦 preview 和正式版输出差异较大,debug 时会归因困难。\n- **版本号语义**:DeepSeek 用日期后缀(0731)代替 v4.1 \u002F v4.2 这种语义版本号,意味着 DeepSeek 在承诺不破 API 兼容性的节奏下追求快速迭代。开源协议下的派生模型如果跟随日期版本号,会很快失去可追踪性。\n\n### 评论\n\n把 V4-Flash 看作开源 LLM 的运维型版本号,比把它当作模型升级更准确。它说明一件事:**开源大模型的差异化竞争,正在从谁的预训练更狠转向谁的后训练管线更稳**。V4-Pro 正式版一旦跟上,这条管线就同时跑通 Flash 和 Pro 两个价位——DeepSeek 在用工程化能力锁住开源 LLM 的商业护城河。\n\n对开发者来说,0731 是一个值得记下的节点:开源 LLM 进入按日期发版的新节奏,你读 release notes 时,要把权重变了和只动了对齐区分开。\n\n### 参考\n\n- DeepSeek 更新日志(2026-07-31):V4-Flash-0731 正式版 API 公测,V4-Pro 正式版将尽快发布\n- MACGPU Blog:DeepSeek V4 GA 把分时电价塞进大模型 API,CSA + HCA 撑起 1M 上下文\n- MorphLLM:DeepSeek V4 混合注意力架构解析——1M 上下文下计算量降至 27%\n- Developers Digest:开源编程模型三选一——GLM-5.2、DeepSeek V4、Qwen3.6","https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3919148874345861","4194681c-1a38-405d-a917-40e1dc2622ea",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"b52db7e9-7c58-42c3-9536-5132cb2f8f72","deepseek",{"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},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"9c177033-332f-4d68-b582-c226ec9421fd","en","DeepSeek-V4-Flash-0731 GA: new post-training, same weights","DeepSeek shipped the V4-Flash-0731 GA API. Architecture, parameter count and weights are identical to the preview release; only post-training was redone. This signals a new rhythm for open-weight LLMs — fast version bumps via RLHF and instruction tuning while the base weights stay frozen. V4-Pro GA is next.","## Post-training only, no new weights: DeepSeek-V4-Flash GA API goes live\n\nOn July 31, DeepSeek posted a release-log update announcing the **DeepSeek-V4-Flash-0731 GA API** is now in public testing, alongside a teaser that V4-Pro GA is coming soon. Compared with V4-Flash-preview, the GA version keeps the **same architecture, parameter count and weights** — only the post-training pass was redone. No new checkpoint, no new architecture, not a single byte of the model file changed.\n\nThis is the textbook move for open-source LLMs that have entered their operational maturity phase: **frozen base weights + iterated post-training**.\n\n### Why \"post-training only\" works\n\nV4-Flash rides DeepSeek's signature stack: **MoE + sparse attention + 1M-token context**. The May V4 GA technical disclosure showed DeepSeek fusing two lines of work — **CSA (Compressed Sparse Attention)** and **HCA (Hierarchical Compressed Attention)** — into a single inference stack, pushing attention compute on a 1M context down to roughly **27%** of dense attention. Peak throughput roughly doubled vs. V3, while API pricing stays aggressively low with a \"time-of-use electricity tariff\" framing.\n\nPost-training cost sits two to three orders of magnitude below pre-training:\n\n- **Pre-training**: thousands of GPUs · tens of days · millions of dollars\n- **Post-training (RLHF \u002F SFT \u002F RL-on-CoT)**: tens to hundreds of GPUs · days to weeks · dramatically lower spend\n\nFlash-preview was named separately because its weights were the by-product of a previous RLHF\u002FSFT pipeline wrap-up — DeepSeek wanted to run one more alignment pass against live traffic. **The 0731 suffix is itself the closing marker of a complete post-training iteration**, not an architectural upgrade.\n\n### Three wins from this playbook\n\n1. **Fast version cadence**: Open-source LLMs can no longer say \"trained from scratch for millions of dollars\" every release. Production users hit regression daily; fixes need to ship in days, not months. A preview-to-GA upgrade that doesn't touch weights can be pushed to all users via API in days.\n\n2. **Behavior calibration without disturbing the base**: V4-Flash is one of the strongest models in its open-weight price band on multilingual, long-context, code and reasoning (see Developers Digest's three-model coding showdown). Those capabilities live in the weights; **changing weights = regression risk**. Changing the alignment layer = changing how the model talks, while preserving base reasoning.\n\n3. **Smooth cost curve**: By the time the GA ships, DeepSeek has stabilized its alignment data, reward model and RL training pipelines. Marginal cost to push out the next post-training iteration collapses. That's why V4-Pro GA can come \"soon\" — the same pipeline swaps cleanly to Pro weights.\n\n### Two open questions for the industry\n\n- **Reproducibility**: From the API surface alone, users can't tell which behaviors come from the weights and which from alignment. When preview and GA diverge, debugging gets murky.\n- **Version semantics**: DeepSeek uses a date suffix (0731) instead of v4.1 \u002F v4.2. They're promising API-compatible fast iteration. Open-source downstream forks that mirror date suffixes will lose traceability fast.\n\n### Take\n\nCalling V4-Flash a \"model upgrade\" is less accurate than calling it an **operational version bump for open-source LLMs**. It tells a clear story: **differentiation in open-source LLMs is shifting from \"who trains hardest\" to \"who runs the steadiest post-training pipeline\"**. Once V4-Pro GA lands, that one pipeline serves both Flash and Pro price points — DeepSeek is using engineering discipline to lock down the commercial moat of open-weight LLMs.\n\nFor developers, 0731 is a marker worth remembering: open-source LLMs have entered \"ship by date\" cadence. When you read release notes, separate \"weights changed\" from \"only alignment changed\".\n\n### References\n\n- DeepSeek release log (2026-07-31): V4-Flash-0731 GA API public test; V4-Pro GA coming soon\n- MACGPU Blog: DeepSeek V4 GA puts time-of-use pricing into the LLM API; CSA + HCA carry a 1M context\n- MorphLLM: DeepSeek V4 hybrid-attention analysis — compute on 1M context down to 27%\n- Developers Digest: Open-source coding models head-to-head — GLM-5.2, DeepSeek V4, Qwen3.6","deepseek-v4-flash-0731-post-training-ga","2026-07-31T10:00:00Z","2026-08-01T00:03:44.258117Z","2026-08-01T00:03:44.258125Z",true,"agent",259,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"d4fa7e14-8fbd-4940-93a6-3dd6f0a3991d","DeepSeek V4 Pro 正式版：1.6T MoE，1M 上下文","deepseek-v4-pro-0813-ga-1m-context-moe","2026-08-13T02:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"31f3215c-0892-419d-a610-fe815cc60bbe","GPT-5.6 降价 80% 把竞争拉进「同等智能成本」：DeepSeek V4 Flash 接招，国产模型卡出双线赛道","gpt-5-6-luna-price-cut-equal-intelligence-cost","2026-08-12T03:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"dbff301b-4dda-4537-8c3f-19ee4a6fd88e","字节跳动正训练 10 万亿参数模型:规模上已与 Anthropic Mythos 5 相当","bytedance-10t-parameter-model-pretraining","2026-08-11T02:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"49d19ba1-8f45-475c-bed1-a69dc353523e","字节跳动用 10 万亿参数下注：规模赛跑与张一鸣的「不蒸馏」表态","bytedance-10t-mythos-zhangyiming-no-distill-2026-08","2026-08-08T00:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"ad10985b-425c-4af1-9495-c63792a2b593","腾讯混元把语音识别打到 3% WER：Hy ASR 3.0 preview 让 ASR 从“逐字”走向“读语境”","tencent-hunyuan-hy-asr-3-0-preview-context-aware","2026-08-05T00:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":64},"d7b6d14d-7257-4794-b92f-31956bbc7eae","原生多模态 vs 后训练加压:国产头部基模两条路线的工程账","native-multimodal-vs-posttraining-2026"]