DeepSeek-V4-Pro 正式版可能已经开始切换。DeepSeek 官方 API 文档悄悄把模型版本从原来的 DeepSeek-V4-Pro 改成 DeepSeek-V4-Pro-0813。调用名称仍是 `deepseek-v4-pro`，上下文为 1M，最大输出 384K，Responses API 现在也已经标为支持。

We raised a $45M Series B led by Peak XV in March, valuing @useblacksmith at $550M. We somehow never got around to announcing it until now; this year has been pretty busy. One thing I’ve been thinking about is how we were both right and wrong about what was coming. When we raised our Series A in April 2025, we believed AI was going to lead to a lot more code being written, and that all of that code would still need to be built, tested, and validated before it could be merged. That part was right, but what we underestimated was how fast it would happen. I have a suspicion that something changed over the holidays. Opus 4.5 came out in November; a lot of developers had time to really use Claude Code over the holidays. I think a lot of people who were still skeptical went back to work in January having fully embraced it. Since the start of the year, the number of CI jobs running on Blacksmith has grown more than 7x. We see customers adopt coding agents, produce way more PRs, and then run into the next problem: how do you actually validate all of this code and get it merged? There are a lot of people working on making code easier to produce. What I find strange is how little attention goes into everything that has to happen after the code is written and before it can be merged. The labs and hyperscalers are mostly focused elsewhere, yet developers and agents spend an enormous fraction of their time here. A lot of companies touch parts of this problem. For us, it’s *the* problem. We started Blacksmith by building a much better CI cloud, but ultimately what we care about is helping developers and agents validate and merge code faster. I sometimes tell candidates that if we don’t solve the whole thing, I’m not sure who will. I don’t think we fully appreciate yet how much more software is going to be produced. The infrastructure we have today wasn’t designed for that volume. There’s a lot we need to rethink from first principles, and I think we’re barely scratching the surface. If you’re extremely good at what you do and this is a problem you want to spend the next several years working on, we’re hiring in SF and NYC:

We've raised a $143M Series C at a $1.5B valuation. In 2023 we spent our time arguing that AI code needed its own reviewer. That argument is over. The work now is moving into the change itself. This is where we think Agentic Change Management comes in.

🚨爆料：DeepSeek Harness 将于今天开启正式公测🔥 刚刚一张据称说DSH 官方内测群的截图在流传 内容是： 0813 计划发布 DSH 公测版，今晚将推送 DeepSeek Harness 最后一个内测版本，让开发者完成最后的插件兼容；插件仓库需要带上 #dsh# topic。 也就是今天，DeepSeek Harness 将正式结束封闭内测，开始对外开放。 同时，内测期间开发的插件也可以迁移到开发者自己的账号下，并正式公开。 看起来 DeepSeek 这次不只是发一个 Harness，连围绕 Harness 的插件生态也准备一起放出来了。 今天可以蹲一下 DeepSeek Harness🫡

Open models may soon be added to an updated AI framework, sources tell WIRED, as the White House continues to grapple with how to regulate a technology it has tried not to regulate.

@cognition Grok 4.7 will exceed all current models. That said, Anthropic is a great company and will probably release improved models soon. However, the SpaceX training corpus is so awesome & unique that I would be shocked if any model is better at real-world engineering than 4.7.

Qwen's biggest open-weight drop yet. 🚀 Meet Qwen3.8-Max: 2.4T params, 95B active, built to take a goal and come back with finished work. ⚡ 1M context. Adjustable reasoning. Parallel tool calls. 💻 Coded unattended for 10+ days, building a self-evolving harness from scratch 🔬 Reproduced a research paper, then ran a 125-hour autonomous loop and beat it 🔧 Took a chip design from RTL to layout, cutting die area by 81%

📄New Research on Self-Evolving Agents: When AI agents modify themselves, how do we know they actually got better? We present Diving into Reliable Self-Evolving Agents: A Survey—a systematic map of how agents self-evolve and what evidence is needed to trust each update. The survey: 🔹 Defines an L0–L4 taxonomy for self-evolving agents 🔹 Introduces a reliability ladder for trustworthy updates 🔹 Curates 549 works in an open companion catalog One core principle: no update should control the only evidence used to accept itself. Explore the full survey ↓ 📄 Paper: 🌐 Project: 💻 GitHub:

腾讯也在财报里披露了一批 AI 业务进展。二季度营销服务收入同比增长 22% 至 436 亿元，腾讯明确把 AI 广告推荐模型和智能投放产品 AIM+ 列为增长动力。营销服务毛利也增长 21% 至 250 亿元。 云业务同样受益于 AI 需求增长。腾讯称，企业服务增长主要由云服务带动，其中 AI 相关服务需求上升是原因之一。 Agent 产品开始出现更明确的付费信号。WorkBuddy 用户快速增长，留存率健康，用户通过订阅和充值购买 Token 的付费意愿强。WorkBuddy 和 CodeBuddy 都实现了「突破性」用户增长，并称目前在各自领域处于中国领先水平。 Hy3 正式版相比 preview 版大幅提升。按 OpenRouter Token 消耗量计算，自 7 月 7 日以来持续位列全球前三。微信 AI Agent「小微」也已开始小范围灰度测试。

腾讯二季度资本开支达到 527.8 亿元，去年同期只有 191.1 亿元，同比增长约 176%。自由现金流则转为负 138 亿元。 腾讯本季经营现金流仍有 527 亿元，但实际支付了 593 亿元资本开支，另有 50 亿元媒体内容付款和 22 亿元租赁负债付款。公司称，现金流里包含大额 AI 算力预付款，用于 Hy 模型升级、WorkBuddy、CodeBuddy、微信 AI 和云服务。剔除这部分预付款，自由现金流仍有 376 亿元。 腾讯的新 AI 产品目前还在明显烧钱。Q2 Non-IFRS 经营利润为 756 亿元；如果不算 Hy、元宝、CodeBuddy、WorkBuddy 和小微，利润会达到 861 亿元。按这一口径粗算，新 AI 产品本季拖累经营利润约 105 亿元。

NVIDIA researchers did it again! They found a way to make KV cache transferable between models. The target model skips prefill entirely, and the conversion runs 2.7 to 25x faster than processing the context again. Let's understand why this is so important today. LLM APIs are stateless, so every turn sends the entire conversation back to the model. The model reads all of it again before writing a single new token, and all of it is billed as input. Prompt caching allows Anthropic and other providers to hold the KV cache for a stable prefix and bill a hit at roughly 10% of the base input rate, because the compute was already done once. The 90% reduction is one of the largest lever in LLM serving, which is why so much production work goes into keeping prefixes byte-stable. But the cache only works on the model that produced it. Keys and values are produced from that model's weights, so no other model can read them. In pratice, the constraint shows up in LLM routing. If the traffic is shifted to a different model for cost/capability reasons, the accumulated KV cache becomes invalid. As a result, the accumulated context has to be processed from scratch, and it's billed at full rate. NVIDIA's recent paper treats this as a representation problem. Prefill's only output is the KV cache, so to move KV between models, we need to convert one model's cache into the format the other expects. They first checked whether the conversion has any structure worth exploiting. They found that moving from Qwen3 14B to 32B, a plain linear regression from a single source layer reconstructed 56% of the variance in the target model's keys. The two models obviously may have different layer counts, so there is no natural one-to-one pairing between them. For each target layer they rank every source layer by how well it predicts that layer, then feed the top eight in together, which takes the reconstruction to 79%. The mapper itself has three parts: > Each target layer and head gets its own independent linear map, solved in one closed-form step rather than by gradient descent. > The cross-layer selection described above is the second part, and their ablation shows it carries the most weight of the three. > Keys also carry a position-dependent rotation from RoPE. They strip that rotation, fit the map in position-free space, then re-apply the target model's rotation at inference. Across six pairs from Qwen3, Llama 3.1 and Ministral 3, four retain 73 to 98% of the receiving model's standalone accuracy, and the conversion runs 3-25x faster than processing the context again. Prior work on cross-model KV reuse exists, but it either trains a neural adapter per pair or requires both models to be architecturally identical. This is probably the first version that is closed-form and training-free, so a lot of it is still open research. Every pair tested belongs to one family, so it works on Qwen to Qwen and Llama to Llama. Cross-family transfer is listed as future work. All six pairs mentioned above also happen to share KV head count and per-head dimension across scales. Mismatched head configurations are currently untested. The researchers scoped this to dense full-attention only, so sliding-window and attention-recurrent hybrids still need work. Here's the paper: Plenty of work is yet to be done. Still, the constraint being solved is genuine. Every model swap currently invalidates the full KV that was already paid for, and this is the first result showing that work might be recoverable without training anything extra. That said, all of this only matters because of what the KV cache is doing in the first place. I wrote a first-principles breakdown of it, covering why the model stores keys and values at all, why the cache grows with every token, and what generation speed looks like with and without it. Read it below.