4 Mac Studios in One Cluster: 1.5 TB of Memory for Trillion-Parameter LLMs

An enthusiast combined four Mac Studio computers into a single cluster with 1.5 TB of shared memory using RDMA over Thunderbolt 5. The cluster runs language models on the scale of DeepSeek V3 and Kimi K2 Thinking, at roughly a trillion parameters. The experiment was published on Habr in July 2026 as a translation of the author's original article.

Why combine four Mac Studios

The four Mac Studios were linked into one cluster to reach 1.5 TB of shared memory, enough to hold a trillion-parameter language model in its entirety. Models of this size physically do not fit into the memory of any single consumer or even server GPU, so the only way to run them locally is to build a memory pool from several machines.

The Mac Studio was chosen for Apple's unified memory architecture: top configurations offer hundreds of gigabytes, accessible to both the CPU and the GPU cores without copying. Four such machines add up to the 1.5 TB total.

• Configuration — four Mac Studio computers in one cluster
• Total memory — 1.5 TB
• Interconnect between nodes — RDMA over Thunderbolt 5
• Target models — DeepSeek V3 and Kimi K2 Thinking (about 1 trillion parameters)
• Performance comparison — against Nvidia and AMD solutions

How the Thunderbolt 5 link works

The cluster nodes are connected over Thunderbolt 5 using RDMA, which gives direct access to a neighboring machine's memory while bypassing the central processor. RDMA removes the main bottleneck of distributed inference: weights and intermediate model states move between nodes with minimal latency and without loading the CPU.

Thunderbolt 5 delivers bandwidth on the order of 80 Gbit/s, and pairing RDMA with such a link is a relatively new approach for the Mac. In effect, the author turned four separate desktops into a single machine with a shared memory address space, distributing the layers of the giant model across them.

What problems came up

The author describes the main pain point as configuration and stability of the four-node setup rather than performance. RDMA over Thunderbolt 5 on macOS is a fresh technology, and the road from "bought four Mac Studios" to "a trillion-parameter model answers reliably" turned out to be far from obvious.

In the article, the author presents speed test results and compares the cluster with Nvidia- and AMD-based solutions, to work out where the Apple setup wins on memory per dollar and where it loses on bandwidth and software maturity.

The goal of the experiment is to run giant language models locally that fit on no existing GPU, as the author describes in the

Habr article.

What this means

Running trillion-parameter models locally is no longer the exclusive preserve of data centers with expensive GPU racks. A cluster of several Mac Studios with unified memory and Thunderbolt 5 is becoming a workable alternative for those who need private inference of large MoE models without the cloud. It is not yet a mainstream solution: rough software and configuration complexity remain a barrier.

Frequently asked questions

Why combine several Mac Studios into a cluster?

To get a single 1.5 TB memory pool. Trillion-parameter models such as DeepSeek V3 and Kimi K2 Thinking do not fit into the memory of a single GPU, while four Mac Studios with shared memory and RDMA over Thunderbolt 5 provide that capacity.

Which models were run locally?

The author ran language models on the scale of DeepSeek V3 and Kimi K2 Thinking on the cluster, both at roughly 1 trillion parameters. The configuration with 1.5 TB of shared memory was built specifically for them.