Desert Ant Labs: local, fast models that run on device

Desert Ant Labs：运行在设备端的本地化、高速模型

Today we’re launching Desert Ant Labs, a European frontier AI lab building opinionated on-device intelligence. We believe the best path to efficient intelligence starts on-device. We’re building small, specialized models for audio, vision, and text – each model answers in milliseconds, and costs nothing to run, so you can put intelligence in every product interaction, without being limited by token cost or inference speed. 今天，我们正式推出 Desert Ant Labs，这是一家位于欧洲的前沿 AI 实验室，致力于构建具有鲜明特色的设备端智能。我们坚信，实现高效智能的最佳路径始于设备端。我们正在为音频、视觉和文本构建小型、专业的模型——每个模型都能在毫秒级时间内响应，且运行成本为零。因此，你可以将智能融入到每一次产品交互中，而无需受限于 Token 成本或推理速度。

Small enough to run on a five-year-old phone, fast enough to use on every frame or keystroke, and better than the API call you’re already paying for. The first 18 models are live today (12 stable and six in beta), accessible via one SDK for Swift, Kotlin, and JavaScript. 这些模型小到可以在五年前的手机上运行，快到足以处理每一帧画面或每一次按键，且性能优于你目前正在付费使用的 API 调用。首批 18 个模型现已上线（12 个稳定版，6 个测试版），可通过适用于 Swift、Kotlin 和 JavaScript 的统一 SDK 进行访问。

One model per task, each built to be the fastest way to complete that task on a device: 每个任务对应一个模型，旨在成为设备上完成该任务的最快方式：

Voz: transcribe 10 minutes of audio in two seconds on an iPhone – 4.7x faster than Whisper – with a start and end time on every word.
Voz： 在 iPhone 上两秒内转录 10 分钟音频——比 Whisper 快 4.7 倍——并提供每个单词的开始和结束时间。
Clear: a 9MB model that can turn a five-minute laptop recording into studio quality audio in one second.
Clear： 一个 9MB 的模型，能在一秒钟内将五分钟的笔记本电脑录音转化为录音室级别的音频。
Redact: mask names, addresses, and card numbers, in real time, in 27 languages, so they never reach your servers.
Redact： 实时屏蔽 27 种语言的姓名、地址和卡号，确保这些敏感信息永远不会到达你的服务器。
Tongue: identify 84 languages from three words, with a 2MB model.
Tongue： 一个 2MB 的模型，仅需三个单词即可识别 84 种语言。

Language ID accuracy, three words in: Tongue (2MB) 0.933 vs. 293MB detector 0.887. Tongue names the language from three words, scoring 0.933 at 2MB against 0.887 for a 293MB detector. And that’s just to name a few. You can find full specs and benchmarks for the other fourteen, on desertant.com/models and Hugging Face. Every model is free up to 100k monthly active devices. No tokens, no logins. 语言识别准确率（三个单词）：Tongue (2MB) 为 0.933，而 293MB 的检测器为 0.887。Tongue 仅凭三个单词即可识别语言，在 2MB 的体积下得分 0.933，而 293MB 的检测器仅为 0.887。这仅仅是其中几个例子。你可以在 desertant.com/models 和 Hugging Face 上找到其余 14 个模型的完整规格和基准测试。每个模型在月活跃设备数 10 万以内均免费。无需 Token，无需登录。

Personal data caught: Redact (12MB) 88.8% vs. GLiNER-PII (2.3GB) 91.1% vs. Rampart (14.7MB) 61.4% vs. OpenAI filter (3GB) 60.2%. Redact catches 88.8% of the personal data in a text, close to the 2.3GB GLiNER-PII, from a 12MB model. 个人数据捕获率：Redact (12MB) 为 88.8%，GLiNER-PII (2.3GB) 为 91.1%，Rampart (14.7MB) 为 61.4%，OpenAI filter (3GB) 为 60.2%。Redact 仅凭 12MB 的模型就能捕获文本中 88.8% 的个人数据，表现接近 2.3GB 的 GLiNER-PII。

We’re building this in Europe, where “on-device” is the sovereign default. The data never leaves your customer’s hands, the feature never depends on someone else’s cloud, and what’s never been uploaded can never be compelled. 我们在欧洲进行开发，在那里，“设备端”是主权默认设置。数据永远不会离开客户手中，功能永远不依赖于他人的云端，而从未上传过的数据也永远不会被强制调取。

How we got here

我们是如何走到这一步的

For five years we’ve been building our video app, Detail, with an on-device first approach. But when we introduced features like Auto Edit to create short clips, or audio enhancement for podcasts, we had to fall back to cloud APIs. And as the popularity of Detail grew, so did our infrastructure bills. Every few months I’d hunt for useful on-device models. I’d surf Hugging Face for a model that could find filler words or clean up a recording. And, every June, we’d get great new tools to build with but the industry wasn’t moving fast enough. 五年来，我们一直以“设备端优先”的理念构建视频应用 Detail。但当我们引入诸如自动剪辑（Auto Edit）生成短片或播客音频增强等功能时，不得不退回到云端 API。随着 Detail 的普及，我们的基础设施账单也随之增加。每隔几个月，我就会寻找有用的设备端模型。我会在 Hugging Face 上搜索能查找填充词或清理录音的模型。每年六月，我们都能获得很棒的新工具，但行业的发展速度还不够快。

The foundation was there: the chips, Core ML, the research. What was missing was everything between that foundation and actually implementing a feature in your app: a model you could drop in and ship with a few lines of code. So, we trained the models ourselves. It turns out training a model is a product design challenge, and product is what we know. 基础已经具备：芯片、Core ML 和相关研究。缺失的是从基础到在应用中实际实现功能之间的所有环节：一个你可以直接放入并用几行代码即可发布使用的模型。因此，我们自己训练了这些模型。事实证明，训练模型是一个产品设计挑战，而产品正是我们所擅长的。

We designed models and local inference that beat cloud services on speed, quality, and cost, and outperform other local and cloud models on the task itself, at a fraction of their size. We replaced Dolby for better, faster audio enhancement with Clear, and made our on-device transcriptions 5x faster with Voz. We also replaced Claude Sonnet with Clips, our 284MB model that turns a 10-minute video into a dozen clips in 5 seconds – 10x faster and using 470x less energy than Sonnet, with the same quality. 我们设计的模型和本地推理在速度、质量和成本上都超越了云服务，并且在特定任务上优于其他本地和云端模型，而体积却只有它们的一小部分。我们用 Clear 取代了 Dolby，实现了更好、更快的音频增强；用 Voz 将设备端转录速度提升了 5 倍。我们还用 Clips 取代了 Claude Sonnet，这是一个 284MB 的模型，能在 5 秒内将 10 分钟的视频转化为十几个片段——速度快了 10 倍，能耗降低了 470 倍，且质量保持不变。

Detail 6, which will launch with iOS 27, replaces all of our cloud APIs with our own models, running entirely on the device. We’ve all spent the past few years building with LLMs as if they were just another API. And, amid the hype around generalist frontier brains, we almost forgot they’re not the only option. 即将随 iOS 27 推出的 Detail 6，将用我们自己的模型取代所有云端 API，完全在设备上运行。过去几年，我们都习惯于将大语言模型（LLM）仅仅当作另一种 API 来使用。在围绕通用前沿大脑的炒作中，我们几乎忘记了它们并非唯一的选择。

Every developer I talk to has a wishlist of on-device models they’d build if cost wasn’t a factor, or a feature they’re bleeding tokens on that they’d happily swap for a local model. A call that runs the same way a hundred thousand times a day: cleaning a recording, tagging a photo, pulling a date out of a sentence, catching a name before the text hits your servers. None of these needs a frontier model. NVIDIA’s own researchers pulled apart three agent systems and estimated that 40 to 70% of their calls to a large model could go to a small, specialized one instead. 我交谈过的每一位开发者都有一个“设备端模型愿望清单”，如果成本不是问题，他们早就构建了；或者他们正因某些功能而消耗大量 Token，如果能换成本地模型，他们会非常乐意。比如每天运行十万次的重复任务：清理录音、标记照片、从句子中提取日期、在文本到达服务器前捕获姓名。这些任务都不需要前沿大模型。NVIDIA 的研究人员拆解了三个智能体系统，估计其中 40% 到 70% 对大模型的调用完全可以由小型专用模型替代。

The compute is already paid for

计算资源早已就绪

The industry will spend about $450 billion on data centers this year. Meanwhile, the world ships more than a billion phones, tablets, and laptops with increasingly capable chips, perfectly suited to these kinds of tasks. There’s more compute available in people’s hands than in every AI data center on earth. We have an unfair advantage with free inference. No per-call cost, so a feature runs on every message instead of the ones you can afford to check. No round-trip, and your customer’s data never leaves the device. When inference costs nothing, the way we build products changes entirely. 今年，整个行业将在数据中心上投入约 4500 亿美元。与此同时，全球每年出货超过 10 亿台手机、平板电脑和笔记本电脑，它们配备了性能日益强大的芯片，非常适合处理这类任务。人们手中掌握的计算能力比地球上所有 AI 数据中心的总和还要多。我们拥有免费推理的非对称优势。没有单次调用成本，因此功能可以在每条消息上运行，而不是仅在你负担得起时才检查。没有往返延迟，客户数据也永远不会离开设备。当推理成本为零时，我们构建产品的方式将发生彻底改变。

Little brains in every product

每个产品中的“小脑”

To build with local models, the developer experience has to get a lot better. You need models you can use commercially, that beat the alternatives on your task in speed and quality, that you can drop into your app with a few lines of code, and are easy to discover. Think of the first hundred models as the cerebellum, the little brain. The little brain handles the always-on work – balance, timing, the skills you never think about, so the rest of the brain is free to think. That’s what we’re building first: fast, specialized models for the work that runs all day, on the device. 要使用本地模型进行开发，开发者体验必须得到大幅提升。你需要的是可以商用、在特定任务上速度和质量优于替代方案、只需几行代码即可集成到应用中且易于发现的模型。将最初的这 100 个模型想象成“小脑”。小脑负责处理那些常驻的工作——平衡、时机、那些你从不刻意去想的技能，这样大脑的其他部分才能腾出空间去思考。这正是我们首先要构建的东西：为全天候运行的任务提供快速、专业的设备端模型。