This week, a developer posted hard numbers on Reddit: a 128GB unified-memory MacBook Pro M5 Max, connected to an AMD Radeon AI PRO R9700 via a Thunderbolt 5 eGPU enclosure (Razer Core X V2), can now run the Qwen3 0.6B–3.8B series of large models locally on macOS — including MoE (Mixture of Experts, which activates parameters on demand for speed) and MTP (Multi-Token Prediction). We note this is not a product launch, but the work of a single open-source developer — he ported the Linux version of the amdgpu driver to macOS.

What this is

In short, Mac users running local large-model inference have historically been limited to Apple Silicon's own GPU (using community tools like llama.cpp / MLX). To use an AMD or Nvidia discrete card, they either moved to a Windows/Linux workstation or waited for Apple's official support — which Apple has never delivered. The new LemonSeed-Engine (LSE) and the ported driver open a third path: AMD GPUs connected externally via Thunderbolt 5/USB4, used as accelerators on macOS.

For developers, this means the M5 Max's 128GB of unified memory can serve as both CPU and GPU memory, paired with an external AMD GPU for inference — the possibility of running 70B-class models locally moves for the first time from "theoretically feasible" to "hands-on testable."

Industry view

Optimistic voices call this "the icebreaker for Apple's local-AI ecosystem," and small-model players long squeezed out by closed hardware may return; the Lemonade SDK organization has gained some traction on GitHub.

But we must flag the reverse risks: first, this is a one-person project with no Apple backing — no signing, no stability guarantees, no enterprise support; second, the Thunderbolt 5 external connection's bandwidth (~80Gbps) is far below internal PCIe, so inference speed will inevitably take a hit; third, Apple Silicon's own Metal accelerators are iterating fast (the M5 Max GPU is already no slouch), and Apple has little incentive to open up third-party GPU access. Once Apple's in-house GPUs catch up, the moat for this approach could simply vanish.

Impact on regular people

For enterprise IT: Deploying "local large models" used to mean provisioning an Nvidia workstation; now there's an additional "Mac workstation + AMD Thunderbolt enclosure" hybrid architecture to consider, but it remains experimental and is not recommended for production environments.

For individual professionals: Short-term impact is limited. If your work machine is already a Mac, you may gain another "offline, no data egress" hardware option for running AI assistants in the next year or two — but the hassle cost of an enclosure plus GPU (typically starting at 15,000 RMB) is nontrivial.

For the consumer market: We won't see a consumer-grade "Mac + AMD GPU" SKU anytime soon; what matters more is the signal it sends — Apple's hardware lockdown is being pried open, one engineering project at a time.