Apple’s new desktop computers are designed specifically for local AI development

苹果新款台式机专为本地 AI 开发而设计

The Mac mini and Mac Studio occupy two distinct points in Apple’s lineup of desktops, but lately, they’ve had something in common: They’re popular for local AI inference and software development thanks to the advantages of their unified memory architecture and the fast CPUs and GPUs on their systems-on-a-chip.

Mac mini 和 Mac Studio 在苹果的台式机产品线中占据着两个不同的定位，但最近它们有了一个共同点：得益于统一内存架构以及片上系统（SoC）中快速 CPU 和 GPU 的优势，它们在本地 AI 推理和软件开发领域变得非常受欢迎。

Today, Apple announced new iterations of both desktops, along with two new chips: the M6, the first 2nm chip in Apple’s M-series lineup for Macs, and the M5 Ultra, now the most powerful chip in the lineup for most things—especially AI workloads. There aren’t any major new features for either machine. This is just a specs bump. But based on how Apple is presenting these refreshes, they’re leaning hard into those use cases, which weren’t even a thought when earlier iterations were first engineered.

今天，苹果发布了这两款台式机的新版本，以及两款新芯片：M6（苹果 Mac M 系列芯片中首款 2nm 芯片）和 M5 Ultra（目前在大多数方面——尤其是 AI 工作负载方面——性能最强大的芯片）。这两款机器都没有重大的新功能，仅仅是规格升级。但从苹果展示这些更新的方式来看，他们正大力向这些使用场景倾斜，而这些场景在早期版本设计时甚至还未被纳入考量。

The devices’ popularity for production inference took off after macOS 26.2 shipped last December. According to Apple’s release notes, 26.2 enabled “low-latency communication between Thunderbolt 5 hosts for use cases including distributed AI inference using MLX.” Thunderbolt 5 is a very fast wired data connection, and MLX is an open source array framework designed to help machine learning workflows take full advantage of the M-series chips’ unified memory.

这些设备在生产环境推理方面的流行始于去年 12 月 macOS 26.2 的发布。根据苹果的发行说明，26.2 版本实现了“Thunderbolt 5 主机之间的低延迟通信，适用于包括使用 MLX 进行分布式 AI 推理在内的用例”。Thunderbolt 5 是一种极速的有线数据连接技术，而 MLX 是一个开源数组框架，旨在帮助机器学习工作流充分利用 M 系列芯片的统一内存。

Since then, both hobbyists and professional developers and researchers have been essentially daisy-chaining Mac minis or Mac Studios to run inference on local large language models that are much bigger than anything that could run a single mass-market device—providing an alternative to ultra-beefy specialized hardware featuring specialized Nvidia GPUs.

从那时起，无论是业余爱好者还是专业开发者和研究人员，都在通过“菊花链”方式连接 Mac mini 或 Mac Studio，以运行比单台大众市场设备所能承载的规模大得多的本地大语言模型推理——这为那些配备专业 Nvidia GPU 的超高性能专用硬件提供了一种替代方案。

The story here is the chips themselves. The M6 is, as expected, a next-generation SoC (system-on-a-chip) meant for a wide range of consumer applications. It has a 12-core CPU, which includes two of what Apple calls “super cores,” alongside four performance cores and six efficiency cores. It’s the first Apple SoC to use all three core types. There aren’t any verifiable benchmarks to work from yet, but Apple claims it’s up to 40 percent faster at multi-threaded CPU performance compared to the M4 two generations ago.

这里的核心在于芯片本身。正如预期，M6 是一款面向广泛消费类应用的新一代 SoC（片上系统）。它拥有 12 核 CPU，其中包括苹果所谓的两个“超级核心”，以及四个性能核心和六个能效核心。这是苹果首款同时使用这三种核心类型的 SoC。目前还没有可验证的基准测试数据，但苹果声称其多线程 CPU 性能比两代前的 M4 快 40%。

It also has a 12-core GPU, which is two cores more than its immediate predecessors, and faster unified memory with up to 160GB per second of bandwidth. Memory capacity is limited to 32GB, though. That’s where the new M5 Ultra—which will be available in the updated Mac Studio—comes in. Its maximum capacity is a whopping 512GB (for the fortunate few who can afford it).

它还配备了 12 核 GPU，比前代产品多出两个核心，并拥有高达 160GB/s 带宽的更快速统一内存。不过，内存容量上限为 32GB。这就是新款 M5 Ultra（将搭载于更新后的 Mac Studio 中）的用武之地。其最大容量高达惊人的 512GB（当然，仅限少数买得起的幸运儿）。

In many respects, it’s literally two M5 Maxes running side by side on one SoC, for 36 CPU cores (12 super, 24 performance) and 80 GPU cores. Apple claims it can achieve up to 1.2TB per second in terms of unified memory bandwidth. Most people would never, ever need that, but that’s what we mean when we say that the AI inference use case is really what Apple is optimizing for here.

在许多方面，它实际上就是将两个 M5 Max 并排运行在同一个 SoC 上，拥有 36 个 CPU 核心（12 个超级核心，24 个性能核心）和 80 个 GPU 核心。苹果声称其统一内存带宽最高可达 1.2TB/s。大多数人永远用不到这种性能，但这就是我们所说的——AI 推理用例确实是苹果此次优化的重点。

That kind of memory and bandwidth is of course useful for other things, like gaming and other 3D graphics applications, but local inference is a growing use case, particularly for software developers. Many developers have been changing their workflows to incorporate AI coding agents more heavily, thanks to powerful frontier large language models and increasingly sophisticated harnesses like Claude Code or Codex, but the cost of running those cloud models is steep, and many are questioning whether it will remain practical.

这种内存和带宽当然也适用于其他用途，如游戏和其他 3D 图形应用，但本地推理是一个正在增长的用例，尤其是对于软件开发者而言。得益于强大的前沿大语言模型以及 Claude Code 或 Codex 等日益成熟的工具，许多开发者正在改变工作流，以更深入地整合 AI 编程助手。然而，运行这些云端模型的成本高昂，许多人开始质疑其是否还能保持实用性。

Open-weight models that can be run locally, like some of the latest Qwen and DeepSeek models, can accomplish many of the same tasks but without charging a fortune for tokens, as they’re smaller and driven by the user’s own local energy and compute. Still, most standard consumer hardware, like an average-spec MacBook Pro, is just far behind enough in terms of the size of models they can run that we’re not fully living in a “just run it all locally on your regular dev workstation” future just yet.

像一些最新的 Qwen 和 DeepSeek 模型这样可以在本地运行的开源权重模型，能够完成许多相同的任务，且无需支付高昂的 Token 费用，因为它们体积更小，且由用户自己的本地能源和算力驱动。尽管如此，大多数标准消费级硬件（如普通配置的 MacBook Pro）在可运行的模型规模上仍有很大差距，这意味着我们尚未完全进入“一切都在常规开发工作站上本地运行”的未来。

That’s why some are tapping these chains of Mac minis or Mac Studios. As noted, that’s the main story with these refreshes, but in other respects they’re standard upgrades for consumers who aren’t interested in all that. Both have Apple’s N1 chip, which supports Wi-Fi 7 and Bluetooth 6. Apple claims the storage is up to twice as fast, at 15GB/s. And this Mac mini ships with 2.5Gb Ethernet as standard, with the option to upgrade to 10Gb.

这就是为什么有些人会利用这些 Mac mini 或 Mac Studio 集群的原因。如前所述，这是此次更新的核心故事，但在其他方面，对于那些对此不感兴趣的消费者来说，它们也只是常规的硬件升级。两者都配备了苹果的 N1 芯片，支持 Wi-Fi 7 和蓝牙 6。苹果声称存储速度提升至原来的两倍，达到 15GB/s。此外，这款 Mac mini 标配 2.5Gb 以太网接口，并可选配升级至 10Gb。

The Mac mini with M6 starts at $899 with 16GB of memory, and configurations with M5 Pro start at $1,699, while the Mac Studio with M5 Max starts at $2,499, and the M5 Ultra configurations start at $5,499. They can, of course, get a lot more expensive than that, depending on how you configure them. Preorders for both start today, and they’re shipping on September 22 (though the 512GB memory config for the Studio won’t ship until late October). Apple also says they’ll ship with macOS 27 Golden Gate, suggesting that the annual OS update may arrive by then for other Mac users, too.

搭载 M6 的 Mac mini 起售价为 899 美元（16GB 内存），M5 Pro 配置起售价为 1,699 美元；搭载 M5 Max 的 Mac Studio 起售价为 2,499 美元，M5 Ultra 配置起售价为 5,499 美元。当然，根据配置的不同，价格可能会高得多。两款产品即日起开启预订，并将于 9 月 22 日发货（不过 Studio 的 512GB 内存版本要到 10 月下旬才会发货）。苹果还表示，这些设备将预装 macOS 27 Golden Gate，这暗示年度操作系统更新届时也可能向其他 Mac 用户推送。