[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepseek-v4-flash-0731-agent-benchmark-official-aug2026":3,"news-related-c94766df-827e-4e4e-a006-b6639ec76722":44},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":30,"news_slug":37,"published_at":38,"created_at":39,"modified_at":40,"is_published":41,"publish_type":42,"image_url":14,"view_count":43},"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 7月31日上线 V4-Flash 正式版 API:284B MoE \u002F 13B 激活的结构、参数量、权重保持不变,仅重新后训练。Terminal-Bench 2.1 跑出 82.7,反超 V4-Pro 预览版的 72.1;同时在 Cybergym、DeepSWE、Toolathlon-Verified 等 Code Agent 基准全面领先。原生支持 Responses API 与 Codex,价格 低(每百万 token 输入)\u002F0.28(输出),MIT 权重。","## 一、Preview 不算数,正式版看的是后训练\n\n2026 年 7 月 31 日,DeepSeek 在 API 文档的更新日志里抛下一条信息密度极高的改动:**DeepSeek-V4-Flash-0731 正式版 API 公测**。调用方式不变,模型名字符串仍然是 `deepseek-v4-flash`,但背后发生的事情和四月那份 preview 完全不同——也完全相同。\n\n完全不同的,是分数。DeepSeek 同步放出的基准对比表显示,0731 这版 checkpoint 在 Code Agent 类任务上的表现全面反超 V4-Pro 预览版:**Terminal-Bench 2.1 82.7**(预览 61.8、Pro 72.1)、**Cybergym 76.7**(预览 38.7、Pro 52.7)、**DeepSWE 54.4**(预览仅 7.3、Pro 12.8)、**Toolathlon-Verified 70.3**(预览 49.7、Pro 55.9)、**DSBench-FullStack 68.7**、**DSBench-Hard 59.6**。这是一个 13B 激活的轻量模型,在 Agent 工作流这种最值得花钱的赛道上,把自家还没转正的旗舰打回了未达预期。\n\n完全相同的,是架构。Hugging Face repo 写得很直白:0731 与 preview 共享同一套权重拓扑,预训练后的 32T tokens 数据不动、284B 总参 \u002F 13B 激活的 MoE 结构不动、1M token 上下文不动、CSA + HCA 混合注意力不动、mHC 残差不动、Muon 优化器不动。**DeepSeek 在不动一分钱新算力、不重训一个 token 的前提下,只靠后训练把模型从「会写代码但不会当 Agent」拉到了「能跑通 Responses 流程、Codex 也能调用」**。开源阵营第一次看到一条清晰的「轻量迭代」信号:权重发布之后,模型还能再变。\n\n## 二、0731 真正变了什么:三条工程流水线\n\nDeepSeek 没有公开后训练的具体配方,但从 release notes、Hugging Face model card 与 vLLM 启动命令三处交叉来看,0731 这一版至少叠了三层 infra 级的能力。\n\n第一层是 **DSpark 推测解码默认装载**。0731 的 checkpoint 直接挂载了 DeepSpec 团队的 DSpark 模块,作为权重的一部分下发;vLLM 跑这条 0731 时只要加一行 `--speculative-config {\\\"method\\\":\\\"dspark\\\",\\\"num_speculative_tokens\\\":7,\\\"draft_sample_method\\\":\\\"greedy\\\"}` 就能启用。DSpark 论文里测的是 60%–85% 的端到端首 token 之后的吞吐提速,且是无损的——这对 Code Agent 这种「每步都要再生成几十 token」的负载,是直接的延迟红利。\n\n第二层是 **Responses API 和 Codex 适配**。0731 是 DeepSeek API 第一次原生支持 OpenAI 的 Responses 形态,且为 Codex 做了专门的工具调用与上下文适配。这意味着开发者把 OpenAI 客户端切到 DeepSeek 后端几乎是「把 base_url 改一下、把 model 字段改成 deepseek-v4-flash」这一笔成本的事——不需要再维护 adapter。\n\n第三层是 **reasoning_effort 的三档可控**。`reasoning_effort` 现在支持 `low`、`high`、`max`。DeepSeek 的推荐采样参数是 temperature=1.0,top_p=0.95(agentic)、1.0(其它),`high` 与 `max` 档位建议最大输出长度 384K tokens。换句话说,0731 可以像 Claude Opus 4 系列那样在「快答模式」与「深推理模式」之间动态切档,且切的是真实训练过的能力,不是 prompt 层的小把戏。\n\n三层叠在一起,0731 的本质是:**同样的预训练、不一样的后训练 + 不一样的 decode 拓扑 + 不一样的 API 形态,带来不一样的 agent 表现**。这是一个把 LLM 当「持续部署服务」而不是「一次性模型发布」来运营的范例。\n\n## 三、价格为什么重要:把 Agent 主线打成基础设施级支出\n\n0731 的定价彻底把这张牌摊明:\n\n- 输入:低 \u002F 百万 token(命中缓存低)\n- 输出:0.28 \u002F 百万 token\n\n作为对比,V4-Pro 的输出价是 0.87 \u002F 百万 token——Flash 只有 Pro 的三分之一。更直接的对比是 Claude Opus 4.8 的 5 \u002F 25 与 OpenAI Codex 的 5 \u002F 30。**0731 把 Terminal-Bench 82.7 这种 Agent 主流基准,做到了 Opus 4.8(85.0)价的 5%**。在 Agent 这条主线上,LLM 调用第一次具备了「按基础设施预算」立项的可能性,而不是「按 SaaS 预算」立项。\n\n## 四、自托管的边界:开源不是万能药\n\n0731 的 MIT 许可证让任何人都能拉下来自己部署,但 304B 参数(含 DSpark 草稿模块)的总量决定了 serving 门槛。DeepSeek 官方给的 vLLM recipe 是单节点 4×GB300;Unsloth 的无量化 dynamic GGUF 在 3-bit 下大约 103 GB,加上 VRAM 共需约 110 GB 内存。这意味着自托管对中小团队并不友好——一台 H100 节点 80GB 显存跑不动 3-bit,4×GB300 整机价格远超普通云年度预算。\n\n但这条路最重要的一扇门**没有关**:开源 + 不可撤回的 MIT 条款 + 官方自带的 serving recipe,意味着任何一家头部云厂商或推理服务商都可以**当天拉权重、当天上线**。事实上 Fireworks 在 7 月 31 日就同步挂出了 serverless endpoint。这是 DeepSeek 把「同等规模模型的发布权」让渡给整个生态的又一次实操——Claude Opus 没法被第三方云以原生模型名卖,DeepSeek V4-Flash 可以。\n\n## 五、需要警惕的两件事\n\n第一,**所有打分都来自 DeepSeek Harness,这个 harness 尚未公开**。DeepSeek 在 benchmark 注释里明确标注:Code Agent 任务使用 minimal mode,max effort,temperature=1.0,top_p=0.95。这个 setting 不可与第三方重跑直接对齐。在独立第三方验证 Cybergym 与 DeepSWE 之前,生产级路由决策不应只基于 DeepSeek 自报数字。\n\n第二,**API 老字符串要废弃了**。`deepseek-chat` 与 `deepseek-reasoner` 在 2026-04-24 的迁移窗口里被临时映射到 V4-Flash 的两种模式,但官方明确这两条字符串将在三个月内下线——也就是 **2026-10-24**。任何继续用这两个字符串的代码,现在开始就该做迁移计划。\n\n## 六、个人评论:这条节奏一旦跑通,价值不在 Flash\n\n0731 真正的影响,不在它本身刷了多少分,而在它示范了一种节奏——**权重发布之后,模型仍然可以以「后训练 + 解码拓扑」的方式不停迭代**。这种节奏放在四月初的 V4-preview 时还不存在,4 个月后,DeepSeek 用一次 0731 把这个节奏立住了。\n\n接下来的 6–12 个月里,这条节奏会被复制到三类玩家身上:\n\n1. **其他中国开源**:Kimi K3 已经把 MoE 训练栈完整开源,具备同样技术条件的团队(阶跃星辰、智谱 GLM-5)会在自家 preview 转正时走同一条路。\n2. **闭源旗舰**:Anthropic 的 Opus \u002F Sonnet 已经在做月度小步快跑,0731 之后闭源厂的「补丁式更新」会被迫更频繁。\n3. **企业自托管党**:Flash 这个尺寸的模型(单节点 4×GB300 可跑)是企业私有化部署最舒服的甜点,V4-Pro 还没转正意味着 Flash 是 DeepSeek 在企业 Agent 市场最拿得出手的牌——也是他们接下来最想守住的一块。\n\n可以预期的是,V4-Pro 正式版一旦上线,同样会沿用这条「后训练驱动」的路线;而 Anthropic 如果要在 V4-Pro + V4-Flash 面前保住 Agent 主线的价位,要么靠更新的产品形态,要么靠更深的工具调用可靠性。\n\n**所以,0731 不只是一次「模型升级」,而是一次「模型何时不再是快照」的范式信号**。开源 LLM 终于把「持续部署」写进了自己的工作流。","https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002FDeepSeek-V4-Flash-0731","24d5c6c5-6573-4180-a1fd-f1459842d1af",[11,15,18,21,24,27],{"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",{"id":28,"name":29,"slug":29,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[31],{"id":32,"lang":33,"title":34,"summary":35,"content":36},"9dbf77db-6172-44db-8bcc-2067884f7494","en","DeepSeek V4-Flash-0731: post-training alone beats V4-Pro on agents","DeepSeek released the V4-Flash official API on July 31: the 284B-MoE \u002F 13B-active architecture, parameter count, and weights are unchanged, only post-training is redone. Across Terminal-Bench, Cybergym, DeepSWE, and Toolathlon-Verified, the model surpasses DeepSeek's own V4-Pro (Preview). Responses API and Codex compatibility included.","## 1. The Preview never counted — the official release is judged on post-training\n\nOn July 31, 2026, DeepSeek dropped a high-information-density change into its API docs update log: **DeepSeek-V4-Flash-0731 official API is now in public beta**. The calling convention does not change — the model string is still  — but what is happening under the hood is, in equal measure, entirely different and entirely the same as the April preview.\n\nWhat is completely different: the scores. DeepSeek's published benchmark table shows that the 0731 checkpoint now surpasses the V4-Pro preview across every Code Agent benchmark the company publishes: **Terminal-Bench 2.1 82.7** (preview 61.8, Pro 72.1), **Cybergym 76.7** (preview 38.7, Pro 52.7), **DeepSWE 54.4** (preview only 7.3, Pro 12.8), **Toolathlon-Verified 70.3** (preview 49.7, Pro 55.9), **DSBench-FullStack 68.7**, **DSBench-Hard 59.6**. This is a 13B-active lightweight model that, on the agent workflows most worth paying for, beat its own un-finalized flagship.\n\nWhat is completely the same: the architecture. The Hugging Face model card is explicit: 0731 shares the exact weight topology with preview. Post-pretraining, 32T tokens of data stay frozen, the 284B total \u002F 13B active MoE structure stays frozen, the 1M-token context window stays frozen, the CSA + HCA hybrid attention stays frozen, the mHC residual stays frozen, the Muon optimizer stays frozen. **Without spending any new compute and without retraining a single token, DeepSeek used post-training to drag a model from \"can write code but does not know how to be an Agent\" to \"can run Responses flows and let Codex call it\".** The open-source camp saw, for the first time, a clear signal of \"lightweight iteration\": after weights are released, a model can still keep changing.\n\n## 2. What actually changed in 0731: three engineering pipelines\n\nDeepSeek did not publicly disclose the post-training recipe, but cross-referencing release notes, the Hugging Face model card, and the vLLM launch command, at least three layers of infra-grade capability were stacked into 0731.\n\nThe first layer is **DSpark speculative decoding shipped by default**. The 0731 checkpoint has DeepSpec's DSpark module mounted as part of the released weights; serving 0731 with vLLM requires only a single  flag. The DSpark paper reports 60–85% end-to-end throughput improvement over the MTP-1 baseline at matched aggregate throughput, and it is lossless. For Code Agent workloads — where every step regenerates dozens of tokens — that is direct latency gain.\n\nThe second layer is **Responses API and Codex adaptation**. 0731 is the first DeepSeek API release with native OpenAI Responses support, and it has been specifically adapted for Codex's tool-calling and context conventions. That means developers can switch an OpenAI client to the DeepSeek backend by changing the base_url and swapping the model string to  — no adapter to maintain.\n\nThe third layer is **a three-tier reasoning_effort switch**.  now supports , , and . DeepSeek's recommended sampling is temperature=1.0, top_p=0.95 (agentic) or 1.0 (other), and  \u002F  modes recommend a maximum output length of 384K tokens. In other words, 0731 can switch between \"fast-answer mode\" and \"deep-reasoning mode\" like Claude Opus 4 does, and the switch toggles real trained-in capability, not a prompt-layer trick.\n\nStacked together, 0731's essence is: **same pretraining, different post-training + different decode topology + different API shape, producing different agent behavior.** This is a paradigm of operating an LLM as a continuously deployed service rather than a one-time model release.\n\n## 3. Why the price matters: turning Agents into infrastructure-grade spend\n\n0731's pricing lays the cards on the table:\n\n- Input: 0.14 USD \u002F million tokens (0.0028 on cache hit)\n- Output: 0.28 USD \u002F million tokens\n\nFor comparison: V4-Pro's output is 0.87 USD \u002F million tokens — Flash costs one-third. The more direct comp is Claude Opus 4.8 at 5 USD \u002F 25 USD and OpenAI Codex at 5 USD \u002F 30 USD. **0731 puts an 82.7 Terminal-Bench Agent-tier score behind a price tag that is 5% of Opus 4.8**. For the first time, an LLM call can be budgeted as infrastructure spend rather than SaaS spend.\n\n## 4. The boundary of self-hosting: open source is not a silver bullet\n\n0731's MIT license lets anyone pull the weights and self-host, but 304B parameters (including the DSpark draft module) sets a clear serving floor. DeepSeek's official vLLM recipe targets a single 4xGB300 node. Unsloth's unquantized dynamic GGUF lands around 103 GB at 3-bit, requiring approximately 110 GB combined RAM + VRAM. That makes self-hosting unfriendly for small and mid-size teams — a single H100 node with 80 GB of VRAM cannot run the 3-bit build, and a full 4xGB300 box is well beyond most annual cloud budgets.\n\nBut the door that matters most is **not closed**: open source + non-revocable MIT + official serving recipe means any major cloud or inference vendor can **pull the weights and ship that day**. Fireworks, in fact, listed a serverless endpoint for 0731 on July 31. This is another instance of DeepSeek ceding \"distribution rights for a same-scale model\" to the ecosystem at large — Claude Opus cannot be sold under its own name by third-party clouds; DeepSeek V4-Flash can.\n\n## 5. Two warnings worth flagging\n\nFirst, **every score comes from DeepSeek Harness, which has not yet been released**. DeepSeek's benchmark notes clearly state: Code Agent tasks were run with minimal mode, max effort, temperature=1.0, top_p=0.95. This setting is not directly comparable to third-party reruns. Until independent third-party verification of Cybergym and DeepSWE lands, production routing decisions should not rely solely on DeepSeek's self-reported numbers.\n\nSecond, **legacy API strings are being deprecated**.  and  were temporarily mapped to V4-Flash's two modes in the April 24 migration window, but DeepSeek has stated these two strings will be retired within three months — i.e. **October 24, 2026**. Any code still using these two strings should start migration planning now.\n\n## 6. Personal take: once this cadence runs, the value is not in Flash itself\n\n0731's real impact is not the score deltas — it is the cadence it demonstrates: **after weights ship, a model can still iterate via post-training plus decode topology as a first-class loop**. That cadence did not exist at V4-preview time in April; four months later DeepSeek has codified it with 0731.\n\nOver the next 6–12 months this cadence will spread to three kinds of players:\n\n1. **Other Chinese open-source labs**: Kimi K3 has already open-sourced its full MoE training stack. Teams with comparable technical depth (StepFun, Zhipu GLM-5) will walk the same path when their preview turns official.\n2. **Closed-source flagships**: Anthropic's Opus \u002F Sonnet already ship monthly small-step updates; after 0731, closed-source vendors' patch-style updates will be forced to be more frequent.\n3. **Enterprise self-hosting**: a model of this size — single-node 4xGB300 is the serving floor — is the sweet spot for enterprise on-prem Agent deployments. Since V4-Pro has not yet gone official, Flash is DeepSeek's most presentable card in the enterprise Agent market — and the one they will most want to defend.\n\nOnce V4-Pro official lands, it will follow the same post-training-driven path; and if Anthropic wants to hold its pricing on the Agent tier against V4-Pro plus V4-Flash, it will need to compete on either newer product shape or deeper tool-call reliability.\n\n**Bottom line: 0731 is not just a model upgrade. It is a paradigm signal that a model is no longer a snapshot. Open-source LLMs have finally written continuous deployment into their own operational loop.**","deepseek-v4-flash-0731-agent-benchmark-official-aug2026","2026-08-01T02:00:00Z","2026-08-01T20:06:42.294354Z","2026-08-01T20:06:42.294363Z",true,"agent",1074,{"items":45},[46,51,56,61,66,71],{"id":47,"title":48,"news_slug":49,"published_at":50},"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":52,"title":53,"news_slug":54,"published_at":55},"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-v4-flash-official-post-training-agent-0731","2026-07-31T08:00:00+00:00",{"id":57,"title":58,"news_slug":59,"published_at":60},"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":62,"title":63,"news_slug":64,"published_at":65},"f9cf9f03-6aca-4d29-94d3-5c6acfeaf435","匿名模型 OX Alpha 短暂登顶 OpenRouter 编码榜:研究者推测底座指向智谱 GLM-5.x","ox-alpha-stealth-openrouter-glm-5-zhipu","2026-08-24T03:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"b4754043-6b19-499f-8459-f8fc786f4d80","Pokee-Isaac 28B 把 10M 上下文塞进客户边界:28B 参数在 RULER 10M 上 93.3%","pokee-isaac-28b-10m-context","2026-08-20T14:00:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"36055e5f-136f-497d-8763-3ed6609f59ff","Meta Muse Glimmer 30B 本地落地:Apache 2.0 的开源智能体,把 Agent 装进 24GB 显存","meta-muse-glimmer-30b-local-agent-apache2-r2","2026-08-19T03:00:00+00:00"]