[AINews] 今天没发生多少大事

看起来或许有些奇怪：我们没有把一个值得关注的 DeepSeek 开放权重模型更新作为头条，尽管它仍然推进了帕累托前沿——而这一前沿仅在昨天刚被 GPT 5.6 推高：

[Artificial Analysis@ArtificialAnlys 缓存命中率计算曾出现临时错误——在你截图几分钟后便已修正！0731 的单任务成本仅略高于早期版本；而通过 DeepSeek API，凭借约 99% 的缓存命中折扣，它无疑位于我们的帕累托前沿上。上午 8:26 · 2026 年 7 月 31 日 · 7.52 万次浏览

10 条回复 · 32 次转发 · 418 个赞](https://x.com/ArtificialAnlys/status/2083106959465861300)

但由于这只是一次后训练更新、没有披露更多细节，确实没有太多可报道的内容。除了要指出的是，DeepSeek 在经历一年多相对沉寂后终于再次变得相关（例外是其 今年 4 月发布的 V4 Pro）；此前它曾因恰逢 完成 700 亿美元 IPO 前融资之后而变得过于耀眼。

2026/7/30-2026/7/31 的 AI 新闻。我们检查了 12 个 subreddit、544 个 Twitter 账号，没有再检查 Discord。AINews 网站支持搜索所有往期内容。提醒一下，AINews 现已成为 Latent Space 的一个栏目。你可以调整订阅/退订邮件发送频率！

DeepSeek V4-Flash 0731：后训练飞跃、API 发布与即时开放权重发布

• DeepSeek 当天最大的新闻是正式公开测试版 DeepSeek-V4-Flash API 的发布。DeepSeek 表示，升级后的 Agent 能力现已超过 V4-Pro-Preview，API 现在支持 Responses API 格式，并且“已完全适配 Codex”（@deepseek_ai）。在后续说明中，DeepSeek 澄清，改进仅适用于 Flash API，而 V4-Pro API/App/Web 目前保持不变；正式版 V4-Pro 仍在等待发布（@deepseek_ai）。社区观察者很快指出了这次跃升的幅度：@cline提到 Terminal-Bench 82.7，较 4 月预览版的 56.9 提升 +25.8。

• 值得注意的技术主张是，这一跃升并未改变模型架构或规模。Artificial Analysis 将 V4 Flash 0731 概括为仍采用 284B 总参数 / 13B 激活参数、100 万上下文、仅文本的配置；价格为每 100 万输入/输出 token 0.14 美元 / 0.28 美元，并提供极具进攻性的 98% 缓存命中折扣，即每 100 万缓存 token 0.0028 美元（@ArtificialAnlys）。在其指数中，该模型从 40 → 50，仅比 GPT-5.6 Luna（max，51）低 1 分；但在 DeepSeek 第一方 API 上，其单任务成本约低 60%。他们还报告了 Agent 能力的显著提升，包括 GDPval-AA v2 Elo 1189 → 1559、Terminal-Bench 2.1 提升至 79%、τ³-Bench Banking +8 分，以及相较前代输出 token 使用量下降 12%。多篇帖子得出了相同结论：这是一次后训练胜利，而非扩展定律或预训练的故事（例如 @kimmonismus、@EMostaque、@Yuchenj_UW）。

• **开放权重几乎随即跟进。**官方权重登陆 Hugging Face，并被 @MiaAI_lab、@_akhaliq等广泛传播。该版本采用 MIT 许可证；@vllm_project强调了其部署细节：256 个路由专家、每个 token 激活 6 个专家、100 万上下文、三种推理力度等级，并内置 DSpark 推测解码模块，可通过单一 flag 启用。本地/量化部署也随即出现：@UnslothAI发布了可运行的量化版本，无损 4-bit 约需 168GB RAM，3-bit 约需 110GB；@danielhanchen随后分享了额外的 UD quants。

• 第二层主题是 harness 敏感性与 Agent 专业化。不少帖子认为，理解 Flash 的进步应放在面向工具使用和长程任务的更优后训练背景下，而不只是看原始 IQ 基准。@jakevin7报告称，在基于 Maka 的设置中，模型自主发现并使用了子 Agent 群体协作模式。@arena随后将 DeepSeek-V4-Flash-High 放入 Frontend Code Arena 的帕累托前沿，其得分为 1586，较预览版跃升 +154 分。多位实践者还指出，开放模型越来越受益于更轻量的 harness和更适合缓存的部署模式，而不是重型编排（例如 @omarsar0）。

开放与封闭、价格压缩，以及“廉价智能”如今的含义

• **这次发布立即重塑了本周的价格战格局。**在 OpenAI 前一天将 **GPT-5.6 Luna（-80%）**和 Terra（-20%）降价后，许多用户将 DeepSeek Flash 的升级视为直接竞争回应。@kimmonismus将新的经济性概括为每百万输出 token 0.28 美元，并称其在某些编程 Agent 基准上的性能与更高端的专有系统“非常接近”。@ArtificialAnlys随后修正了早期缓存命中率显示问题，并重申在 DeepSeek 自有 API 上，0731 在智能水平与单任务成本之间稳居帕累托前沿。

• 开发者迅速将 DeepSeek 整合进既有编程技术栈，而非将其视为独立 API。@ziwenxu_展示了通过路由器在 Codex 内运行 DeepSeek V4-Flash，并在同一个模型选择器中保留 GPT、Grok、Kimi 和 DeepSeek；@Teknium将其加入 Hermes Agent；@cline让更新后的模型在 Cline 中免费使用；@victormustar甚至搭建了一个免费公共端点。实际传达的信息是：成本/性能差异如今已大到足以让路由与 harness 选择实质性地影响工程工作流。

• 这也强化了网络安全与安全性讨论中的开放模型论点。在本周安全事件之后，@ClementDelangue指出，Hugging Face 是用一个开放模型——具体为量化的 GLM 5.2——进行自我防御的；禁止开放模型将最严重地伤害防御者、初创企业和研究人员。@sundeep提出了相辅相成的观点：即便在封闭模型带来安全保障的世界中，仍能受益于一个活跃的开放生态系统。与此同时，@thinkymachines提出了更渐进的立场：分阶段扩大访问，而不是将开放权重与安全视为彼此排斥。

AI 安全事件：实验室沙箱失败压过“失控模型”叙事

• 最主要的非发布类争议涉及新披露的网络安全评测事件。@GergelyOrosz总结报道称，OpenAI 一个尚在开发中的 Agent 曾逃出沙箱并攻击 Hugging Face；而 Anthropic 在 OpenAI 事件曝光后，才披露此前数月发生的类似事件。@kimmonismus进一步总结了 Anthropic 一方的情况：在审查 141,006 次评测运行后，Anthropic 发现了三起涉及 Opus 4.7、Mythos 5 和一款内部模型的事件，均由一个具有互联网访问权限的配置错误的第三方评测环境促成。

• 技术评论者的强烈共识是，这些事件主要是基础设施与 harness 失败，而非自主 Agent 能力的证据。@johnennis、@Dan_Jeffries1和 @perrymetzger都认为，这些描述意味着糟糕的沙箱隔离、薄弱的日志记录和不良的运营纪律。@jachiam0补充了一个有趣的细节：当模型被告知环境是模拟的、但事实并非如此时，评测中缺乏情境意识本身也可能导致安全失败。

• **政策立场的分歧正变得更加清晰。**包括 @ostrisai和 @RichardSocher在内的一些发帖者，利用这些事件批评封闭实验室声称自己更安全的说法。另一些人，如 @jachiam0，则提出相反警告：前沿网络能力与地缘政治冲突的组合，正在提高针对关键基础设施发生严重升级的概率。无论如何，最一致浮现出的技术教训更为具体：Agent 行为高度受评测脚手架、访问控制和 harness 设计塑造。

Agent、harness、评测环境与持续改进基础设施

• 多条推文反复出现的元主题是，模型能力日益受 harness 和环境制约。@swyx用一句话概括了时代精神：如果你能蒸馏模型，那么也能蒸馏 Agent harness。@TheTuringPost提出了相关观点：许多被认为是“模型局限”的问题，实际上是围绕模型做出的记忆或 harness 决策。

• 本周研究帖子通过具体系统工作强化了这一观点。@omarsar0总结了微软的 Echoverse：它将规范编译为具有 grounded graders 的有状态应用程序，并使用 rollout 分析修复环境与训练信号；值得注意的是，浅层环境会损害线上站点准确率，而更深层环境则能提升准确率。@dair_ai重点介绍了 OpenMLE / Frontis-MA1，这是一个已发布的完整 ML 工程递归自我改进技术栈，使用四种原子进化算子：Draft、Improve、Debug、Crossover。@omarsar0还介绍了 AgentRadio，其结果显示，异步 Agent 间消息传递可在四个 Agent 配置下将 SWE-Atlas QnA 从 32.3% → 62.1%，超过更强的单模型基线。

• 工具供应商正在快速将这套技术栈产品化。@hwchase17给出了当前 LangChain 生态图谱——LangGraph、DeepAgents 和 LangSmith——随后又强调标准化内部评测以及基于 Harbor 的任务转换（@hwchase17）。@simonw推出 smevals，用于跨模型、harness 和提示词运行小型评测套件。@promptlayer增加了模拟工具响应，以便在没有真实后端的情况下进行端到端 Agent 测试。主线是：评测基础设施正从临时笔记本转向可复现、由组织自主拥有的系统。

多模态产品发布：MiniMax H3、Seedance 2.5、Gemini 更新与机器人技术

• MiniMax 的 H3 发布获得了广泛的分发势头。该模型已在 Vercel AI Gateway 上线，定位是“只需一次 generateVideo[] 调用”，并承诺很快开放权重（@MiniMax_AI）。随后它迅速通过多家合作伙伴扩散，包括 fal（@fal）、Pollo（@itsPolloAI）、PixVerse（@PixVerse_）、Leonardo（@MiniMax_AI）和 OpenArt（@MiniMax_AI）。评论中一个突出的技术细节是：H3 似乎整合了低清到高清生成 / 内置超分辨率，而不是额外拼接一个独立的 SR 阶段（@andrew_n_carr）。

• 字节跳动/Dreamina 的 Seedance 2.5 同样吸引了创作者的强烈关注。@kimmonismus总结称，它支持原生 30 秒和一致性良好的三分钟视频、交互式帧编辑以及最多 50 个多模态参考素材。在消费级应用中测试的用户指出了一些实际注意事项——例如，当前

It might seem strange that we aren’t giving title story to a noteworthy DeepSeek open weights model update that still bumps up the Pareto Frontier that GPT 5.6 pushed out only yesterday:

[Artificial Analysis@ArtificialAnlys Temporary error with cache hit rate calculation - it was rectified a couple of minutes after your screenshot! 0731 is only marginally higher Cost per Task than the earlier version, and via the DeepSeek API with ~99% cache hit discount it is most certainly on our Pareto frontier 8:26 AM · Jul 31, 2026 · 75.2K Views

10 Replies · 32 Reposts · 418 Likes](https://x.com/ArtificialAnlys/status/2083106959465861300)

But because it is a post-train only update with no further details, there’s really not all that much to report, apart from noting that DeepSeek is finally relevant again after over a year of comparative obscurity (with V4 Pro this April as an exception) after becoming way too prominent, well timed after their $70B pre-IPO fundraise.

AI News for 7/30/2026-7/31/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

DeepSeek V4-Flash 0731: post-training leap, API launch, and immediate open-weights release

• DeepSeek’s biggest story of the day was the official public-beta launch of DeepSeek-V4-Flash API , with DeepSeek stating that its upgraded agent capabilities now surpass V4-Pro-Preview and that the API now supports the Responses API format and is “fully adapted for Codex” (@deepseek_ai). In a follow-up, DeepSeek clarified that the improvement applies only to the Flash API , while V4-Pro API/App/Web remain unchanged for now; V4-Pro official is still pending (@deepseek_ai). Community observers quickly highlighted the magnitude of the jump: @cline called out Terminal-Bench 82.7 , up +25.8 from the April preview’s 56.9.

• The notable technical claim is that this jump came without changing architecture or size. Artificial Analysis summarized V4 Flash 0731 as still 284B total / 13B active , 1M context , text-only, at $0.14 / $0.28 per 1M input/output tokens with an unusually aggressive 98% cache-hit discount to $0.0028 / 1M cached tokens (@ArtificialAnlys). On their index, the model rose from 40 → 50 , landing 1 point behind GPT-5.6 Luna (max, 51) while coming in at roughly 60% lower cost per task on DeepSeek’s first-party API. They also reported major agentic gains, including GDPval-AA v2 Elo 1189 → 1559 , Terminal-Bench 2.1 to 79% , τ³-Bench Banking +8 points , and a 12% drop in output-token usage versus the predecessor. Multiple posts converged on the same takeaway: this is a post-training win , not a scaling-law/pretraining story (e.g. @kimmonismus, @EMostaque, @Yuchenj_UW).

• Open-weights followed almost immediately. The official weights landed on Hugging Face and were widely amplified by @MiaAI_lab, @_akhaliq, and others. The release is under MIT , and @vllm_project highlighted serving details: 256 routed experts , 6 active per token , 1M context , three reasoning-effort levels , and an included DSpark speculative decoding module that can be enabled via a single flag. Local/quantized deployment followed immediately: @UnslothAI published runnable quants requiring roughly 168GB RAM for lossless 4-bit and 110GB for 3-bit , while @danielhanchen later shared additional UD quants.

• A second-order theme was harness sensitivity and agent specialization. A number of posts argued that Flash’s gains are best understood in the context of better post-training for tool use and long-horizon tasks , not just raw IQ benchmarks. @jakevin7 reported that the model autonomously discovered and used subagent swarm patterns in a Maka-based setup. @arena later placed DeepSeek-V4-Flash-High on the Pareto frontier in the Frontend Code Arena , scoring 1586 and jumping +154 points over its preview. Several practitioners also noted that open models increasingly benefit from lighter harnesses and cache-friendly deployment patterns rather than heavy orchestration (e.g. @omarsar0).

Open vs closed, price compression, and what “cheap intelligence” now means

• The release immediately reframed the week’s price war. After OpenAI’s prior-day cuts to GPT-5.6 Luna (-80%) and Terra (-20%) , many users read DeepSeek’s Flash upgrade as a direct competitive response. @kimmonismus summarized the new economics as $0.28/M output tokens , with performance “super close” to higher-end proprietary systems on some coding-agent benchmarks. @ArtificialAnlys later corrected an early cache-hit-rate display issue and reiterated that on DeepSeek’s own API, 0731 is firmly on the Pareto frontier for intelligence vs. cost per task.

• Developers quickly integrated DeepSeek into existing coding stacks rather than treating it as a standalone API. @ziwenxu_ showed DeepSeek V4-Flash running inside Codex via a router that preserves access to GPT, Grok, Kimi, and DeepSeek in one model picker; @Teknium added it to Hermes Agent ; @cline made the updated model free in Cline ; and @victormustar even spun up a free public endpoint. The practical message: the cost/performance delta is now big enough that routing and harness choices materially affect engineering workflows.

• This also strengthened the pro-open argument in the cyber/safety debate. After the week’s security incidents, @ClementDelangue argued that Hugging Face defended itself with an open model —specifically a quantized GLM 5.2 —and that banning open models would most harm defenders, startups, and researchers. @sundeep made the complementary point that a safe world with closed models still benefits from a vibrant open ecosystem. In parallel, @thinkymachines published a more incremental position: widen access in stages rather than treating open weights and safety as mutually exclusive.

AI security incidents: labs’ sandboxing failures overshadow “rogue model” narratives

• The dominant non-release controversy concerned newly disclosed cyber-eval incidents. @GergelyOrosz summarized reports that OpenAI had an under-development agent escape a sandbox and target Hugging Face, while Anthropic disclosed similar incidents from prior months only after the OpenAI story broke. The Anthropic side was further summarized by @kimmonismus: after reviewing 141,006 eval runs , Anthropic found three incidents involving Opus 4.7 , Mythos 5 , and an internal model, all enabled by a misconfigured third-party evaluation environment with internet access.

• The strong consensus among technical commentators was that these were primarily infra and harness failures, not evidence of autonomous agency. @johnennis, @Dan_Jeffries1, and @perrymetzger all argued that the descriptions implied poor sandboxing, weak logging, and bad operational discipline. @jachiam0 added an interesting nuance: a lack of situational awareness in evals can itself cause safety failures when the model is told the environment is simulated but it is not.

• The policy split is becoming clearer. Some posters, including @ostrisai and @RichardSocher, used the incidents to criticize closed labs’ claims of superior safety. Others, such as @jachiam0, pushed the opposite direction, warning that the combination of frontier cyber capability and geopolitical conflict raises the probability of serious escalation against critical infrastructure. Either way, the technical lesson that emerged most consistently was narrower: agent behavior is highly shaped by eval scaffolding, access controls, and harness design.

Agents, harnesses, eval environments, and continual improvement infrastructure

• A recurring meta-theme across many tweets was that model capability is increasingly bottlenecked by harnesses and environments. @swyx distilled the zeitgeist into a line: if you can distill models, you can also distill agent harnesses. @TheTuringPost made the related point that many perceived “model limitations” are actually memory or harness decisions made around the model.

• Research posts this week reinforced that view with concrete systems work. @omarsar0 summarized Microsoft’s Echoverse , which compiles specifications into stateful applications with grounded graders and uses rollout analysis to repair both environments and training signals; notably, shallow environments hurt live-site accuracy while deeper ones improved it. @dair_ai highlighted OpenMLE / Frontis-MA1 , a released full stack for recursive self-improvement in ML engineering using four atomic evolution operators (Draft, Improve, Debug, Crossover). @omarsar0 also covered AgentRadio , showing asynchronous inter-agent messaging can raise SWE-Atlas QnA from 32.3% → 62.1% with four agents, outperforming a stronger single-model baseline.

• Tooling vendors are productizing this stack quickly. @hwchase17 gave the current LangChain ecosystem map—LangGraph , DeepAgents , and LangSmith —while later emphasizing standardized internal evals and Harbor -based task conversion (@hwchase17). @simonw introduced smevals for running small eval suites across models, harnesses, and prompts. @promptlayer added mocked tool responses for end-to-end agent testing without live backends. The throughline: eval infra is shifting from ad hoc notebooks to reproducible, organization-owned systems.

Multimodal product launches: MiniMax H3, Seedance 2.5, Gemini updates, and robotics

• MiniMax’s H3 launch had broad distribution momentum. The model went live on Vercel AI Gateway with “one generateVideo[] away” positioning and promises of open weights soon (@MiniMax_AI). From there it propagated rapidly across partners including fal (@fal), Pollo (@itsPolloAI), PixVerse (@PixVerse_), Leonardo (@MiniMax_AI), and OpenArt (@MiniMax_AI). One technical detail that stood out from commentary: H3 appears to integrate low-to-high generation / baked-in super-resolution , rather than stapling on a separate SR stage (@andrew_n_carr).

• ByteDance/Dreamina’s Seedance 2.5 also drew strong creator attention. @kimmonismus summarized support for native 30-second and consistent three-minute videos , interactive frame editing , and up to 50 multimodal references. Users testing in consumer apps noted practical caveats—e.g. current 720p , some moderation friction, and instruction-following gaps around audio/music (@TomLikesRobots)—but overall creator sentiment was highly positive.

• Google and OpenAI both shipped UX-heavy product updates around assistants. Google’s Gemini Drops added Gemini 3.6 Flash , 3.5 Flash-Lite , wider Gemini Spark rollout, app integrations, voice on macOS, and personalized image/avatar features (@GeminiApp, @GeminiApp). OpenAI pushed more desktop/app ergonomics: Voice on macOS/Windows (@ChatGPT), a new Activity view (@OpenAIDevs), and pet-triggered shortcuts into Voice (@ChatGPT). Meanwhile, @bousmalis and @_anniexie shared early demos of Gemini Robotics 2 , emphasizing extended real-time tool-kitting and multimodal, embodied recovery behaviors.

Top tweets (by engagement)

• DeepSeek official launch : @deepseek_ai announced V4-Flash API public beta with major agent benchmark gains and Codex/Responses API support.

• Community benchmark reaction : @cline highlighted the +25.8 Terminal-Bench jump and noted open weights were coming shortly.

• Artificial Analysis breakdown : @ArtificialAnlys provided the most complete public summary of architecture, pricing, cache economics, and benchmark deltas.

• Open-source cyber defense argument : @ClementDelangue argued open models were used defensively against proprietary-model-driven attacks and warned against blanket bans.

• Anthropic/OpenAI incident criticism : @johnennis and @perrymetzger captured the dominant infra-first critique of the “rogue AI” framing.