[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepseek-v4-flash-official-post-training-agent-0731":3,"news-related-3cc63477-1334-497d-80cb-90850c019101":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},"3cc63477-1334-497d-80cb-90850c019101","DeepSeek-V4-Flash 转正:不靠换架构,只做后训练重新发力 Agent","DeepSeek 7月31日上线 V4-Flash 正式版 API,模型结构与尺寸与 preview 完全一致,仅重新做后训练。在 Code Agent 类基准上大幅超越 V4-Pro-Preview,Terminal Bench 2.1 达 82.7,DSBench-FullStack 68.7。同步宣布 V4-Pro 正式版即将发布。","# DeepSeek-V4-Flash 转正:不靠换架构,只做后训练重新发力 Agent\n\n## 一、Preview 不算数,正式版看后训练\n\n7 月 31 日,DeepSeek 在官方 API 文档的 Update Log 投下一条不算显眼但信息密度极高的更新:**DeepSeek-V4-Flash 正式版 API 公测上线**。调用方式不变,模型名直接设为 `deepseek-v4-flash` 即可。\n\n这条更新最关键的反直觉之处在于:DeepSeek 明确声明,**V4-Flash-0731 的模型结构、尺寸与 DeepSeek-V4-Flash-Preview 完全一致**。换言之,这次\"转正\"既没有换架构,也没有扩参数,只是重新做了一遍后训练(post-training)。在动辄几万亿参数、几百万卡时的 LLM 行业里,这种\"动训练不动架构\"的迭代方式并不常见——它说明 V4-Flash 的底座已经被认为足够好,真正缺的是面向 Agent 场景的对齐。\n\n## 二、Agent 基准大涨,Codex 适配是另一条主线\n\n伴随这次更新,DeepSeek 同步放出了一组针对 Agent 能力的 benchmark:\n\n- Terminal Bench 2.1:**82.7**\n- NL2Repo:**54.2**\n- Cybergym:**76.7**\n- DeepSWE:**54.4**\n- Toolathlon verified:**70.3**\n- Agent Last Exam:**25.2**\n- Automation Bench (Public):**25.1**\n- DSBench-FullStack:**68.7**(内部全栈开发测试集)\n- DSBench-Hard:**59.6**(内部 Coding Agent 难题测试集)\n\n官方说法是\"显著增强的 agent 能力,benchmark 结果远超 V4-Pro-Preview\"。这个对比对象挑得很有意思——既然 V4-Flash 在结构上是 Flash 体量,却能在 Code Agent 类任务上压过体量更大的 Pro-Preview,说明后训练阶段他们把大量精力压在了 Agent 路径上,而非传统的 NLP 通用能力。\n\n测试条件也披露了:Code Agent 类任务用 DeepSeek 自家 Harness 最小模式(即将开源)做框架,effort 拉满,top_p=0.95、temperature=1.0。这意味着分数带一定的\"上限运行\"色彩——但反过来,对调用方也是好消息,只要给足 budget,Flash 也能跑出 Pro 级别表现。\n\n另一个工程层面的细节是:**V4-Flash 正式版原生支持 Responses API 格式,并针对 Codex 做了专门适配**。配置文档已挂在 `api-docs.deepseek.com\u002Fquick_start\u002Fagent_integrations\u002Fcodex`。这等于把 DeepSeek 模型摆到了 OpenAI Codex 客户端面前——对企业用户来说,迁移成本进一步下降。\n\n## 三、不动 V4-Pro 与 Web 端,V4-Pro 正式版\"很快\"\n\n官方刻意强调:**本次只升级 V4-Flash 的 API,V4-Pro API 与 APP\u002FWEB 模型均保持不变**。这一句值得玩味:Flash 已经有了能压过 Pro-Preview 的 Agent 分数,那 Pro 正式版还有什么可发的?\n\nDeepSeek 给出的答案是\"V4-Pro 正式版将会尽快发布\"。考虑到 V4-Flash 在 Code Agent 上的表现已经超过 V4-Pro-Preview,Pro 正式版大概率不是\"Flash 的放大版\",而是会在长上下文、复杂推理、多步规划上做出 Pro 级别的差异化定位。换句话说,DeepSeek 正在用 Flash 打 Code Agent 这一最卷、最容易看出\"性价比\"的战场,而把 Pro 留给真正需要深度推理的科研\u002F金融\u002FAgent 长链场景。\n\n## 四、所以呢?\n\n对开发者来说,这次更新最大的实际意义是:**V4-Flash 已经具备作为 Coding Agent 默认底座的能力**。Terminal Bench 82.7、DSBench-FullStack 68.7 这个数字组合,在开源 MoE 阵营里相当能打。对中小团队而言,意味着可以以更低成本跑类似 Codex 风格的工作流,不必死磕 GPT-5 或 Claude Opus。\n\n对行业来说,DeepSeek 用行动验证了一件大家心里隐约知道、但少有人敢公开声明的事:**模型能力的天花板,后训练阶段还有相当大的挖掘空间**。只要底座规模够,后训练可以在不增加推理成本的前提下,把 Agent 能力再拉一档。这对所有\"小模型 + 强后训练\"的路线都是好消息。\n\n接下来值得盯的两件事:**V4-Pro 正式版的差异化点是什么**;以及 DeepSeek Harness 最小模式开源后,会不会带动一波国产 Agent 框架的标准化。\n\n**参考来源**\n- DeepSeek API Update Log(2026-07-31):https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews260731","https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews260731","4194681c-1a38-405d-a917-40e1dc2622ea",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b52db7e9-7c58-42c3-9536-5132cb2f8f72","deepseek",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":25,"name":26,"slug":26,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"8300509d-5465-4fed-9fbf-3f31434f6670","en","DeepSeek-V4-Flash GA: post-training, not new architecture","On July 31 DeepSeek shipped the official V4-Flash API — same architecture and size as the preview, only post-training was redone. Agent benchmarks jump sharply over V4-Pro-Preview: Terminal Bench 2.1 82.7, DSBench-FullStack 68.7. V4-Pro official is promised \"soon.\"","# DeepSeek-V4-Flash Goes Official: No Architecture Change, Post-Training Does the Heavy Lifting for Agents\n\n## 1. Preview Wasn't Real — the Official Release Lives in Post-Training\n\nOn July 31, DeepSeek dropped a low-key but information-dense update in the official API documentation Update Log: **the DeepSeek-V4-Flash official API is now in public beta**. The calling convention is unchanged — just set the model name to `deepseek-v4-flash`.\n\nThe most counter-intuitive part of this update is that DeepSeek explicitly states **V4-Flash-0731 keeps the exact same model architecture and size as the DeepSeek-V4-Flash preview**. In other words, this \"promotion to official\" did not swap architectures or expand parameters — only a fresh round of post-training. In an industry accustomed to tens-of-trillion parameters and millions of GPU-hours, this kind of \"train harder, don't architect differently\" iteration is rare. It signals that the V4-Flash base is already considered good enough, and what was actually missing was Agent-scenario alignment.\n\n## 2. Agent Benchmarks Surge, and the Codex Adaptation Is a Second Headline\n\nAlong with the release, DeepSeek published a batch of Agent-focused benchmark numbers:\n\n- Terminal Bench 2.1: **82.7**\n- NL2Repo: **54.2**\n- Cybergym: **76.7**\n- DeepSWE: **54.4**\n- Toolathlon verified: **70.3**\n- Agent Last Exam: **25.2**\n- Automation Bench (Public): **25.1**\n- DSBench-FullStack: **68.7** (internal full-stack dev test set)\n- DSBench-Hard: **59.6** (internal Coding Agent hard-problem test set)\n\nDeepSeek's framing: \"Significantly enhanced agent capabilities, benchmark results far exceeding V4-Pro-Preview.\" The choice of comparison is telling — if a Flash-tier model beats a Pro-tier preview on Code Agent tasks, the post-training team clearly poured energy into the Agent path rather than conventional NLP generality.\n\nTest conditions were also disclosed: Code Agent tasks ran under DeepSeek's own Harness minimal mode (open-sourcing soon), with effort maxed out, top_p=0.95, temperature=1.0. That tints the scores with \"ceiling-run\" energy — but on the flip side, it's good news for callers: give Flash enough budget and it'll produce Pro-grade output.\n\nAnother engineering footnote: **V4-Flash official natively supports the Responses API format, with dedicated adaptation for Codex**. The configuration doc is now live at `api-docs.deepseek.com\u002Fquick_start\u002Fagent_integrations\u002Fcodex`. This effectively puts DeepSeek models in front of OpenAI Codex clients — for enterprise users, the migration cost drops another notch.\n\n## 3. Don't Touch V4-Pro or Web Yet — the V4-Pro Official Is \"Soon\"\n\nThe official wording is deliberate: **this update only upgrades the V4-Flash API; V4-Pro API and APP\u002FWEB models remain unchanged**. That's worth reading carefully. If Flash already beats Pro-Preview on Agent scores, what's left for the Pro official release?\n\nDeepSeek's answer: \"The official release of DeepSeek-V4-Pro will follow soon.\" Given that V4-Flash already outscores V4-Pro-Preview on Code Agent, the Pro official is unlikely to be \"a scaled-up Flash.\" It will probably differentiate on long context, deep reasoning, and multi-step planning — i.e. Pro-tier positioning for research, finance, and long Agent chains. Translation: DeepSeek is putting Flash in the ring for the most contested, most price-sensitive battlefield (Code Agents), while reserving Pro for genuinely deep-reasoning scenarios.\n\n## 4. So What?\n\nFor developers, the practical takeaway: **V4-Flash is now strong enough to be a default Coding Agent backbone**. A Terminal Bench 82.7 \u002F DSBench-FullStack 68.7 combination is genuinely competitive within the open-source MoE camp. Small and mid-sized teams can run Codex-style workflows at lower cost without having to push everything onto GPT-5 or Claude Opus.\n\nFor the industry, DeepSeek quietly validated something many suspected but few dared to state publicly: **the ceiling of model capability still has plenty of room in post-training**. As long as the base is large enough, post-training can pull Agent capability up another notch without raising inference cost. That's good news for every \"small model + heavy post-training\" path.\n\nTwo things worth watching next: **what the V4-Pro official actually differentiates on**, and whether the open-sourcing of DeepSeek's Harness minimal mode kicks off a wave of standardization in domestic Agent frameworks.\n\n**Source**\n- DeepSeek API Update Log (2026-07-31): https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews260731","deepseek-v4-flash-official-post-training-agent-0731","2026-07-31T08:00:00Z","2026-07-31T18:03:46.800620Z","2026-07-31T18:03:46.800634Z",true,"agent",185,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"c94766df-827e-4e4e-a006-b6639ec76722","DeepSeek V4-Flash-0731 转正观察:权重不动,后训练把 Agent 分数打到 V4-Pro 之上","deepseek-v4-flash-0731-agent-benchmark-official-aug2026","2026-08-01T02:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"c4ec4625-4a84-4c24-88f0-0ef1beb4f19e","Grok 4.6 发布:61 分追平 GPT-5.6 Sol,把长程 Agent 的 token 账单砍到四分之一","grok-4-6-agentic-cost-frontier","2026-08-14T19:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"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":59,"title":60,"news_slug":61,"published_at":62},"fbdcfd41-bb54-487a-8319-9f35dfc82be5","DeepSeek-V4-Flash转正：这次升级不靠换架构","deepseek-v4-flash-0731-agent-update","2026-07-31T08:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"e75069c6-f15c-4ff9-8b11-404d705442e8","Upstage Solar Pro 4:把「agent 跑得稳」做成新一代闭源模型卖点","upstage-solar-pro-4-agent-reliability-closed-llm","2026-08-25T03:00:00+00:00"]