4 Mac Studio в один кластер: 1,5 ТБ памяти для LLM на триллион параметров

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

Why combine four Mac Studios

Four Mac Studios were linked into a single cluster to get 1.5 TB of shared memory — a volume large enough to hold an entire trillion-parameter language model. Models of this size physically don't fit in the memory of any single consumer or even server-grade graphics card, so the only way to run them locally is to pool memory from several machines.

Mac Studio was chosen because of Apple's unified memory architecture: top-tier configurations offer hundreds of gigabytes, accessible to both the CPU and the GPU core without copying. Four such machines together yield a combined 1.5 TB.

*Configuration — four Mac Studio computers in one cluster

*Shared memory — 1.5 TB

*Inter-node link — RDMA over Thunderbolt 5

*Target models — DeepSeek V3 and Kimi K2 Thinking (around 1 trillion parameters)

*Performance comparison — against Nvidia and AMD solutions

How the Thunderbolt 5 link works

The cluster's nodes are connected over Thunderbolt 5 using RDMA — direct access to a neighboring machine's memory that bypasses the central processor. RDMA is exactly what removes the main bottleneck of distributed inference: exchanging weights and intermediate model states between nodes happens with minimal latency, without loading the CPU.

Thunderbolt 5 delivers bandwidth of roughly 80 Gbit/s, and pairing RDMA with such a channel is a relatively new approach for Mac. In effect, the author turned four separate desktops into one 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 not as performance, but as the setup and stability of linking four nodes together. RDMA over Thunderbolt 5 on macOS is a fresh technology, and the path from "bought four Mac Studios" to "the trillion-parameter model responds stably" turned out to be far from obvious.

In the article, the author presents speed test results and compares the cluster to Nvidia- and AMD-based solutions — to see 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 locally giant language models that don't fit on any single existing graphics card — as the author describes in the article on

Habr.

What this means

Running trillion-parameter models locally is no longer the exclusive domain of data centers with expensive GPU racks: a setup of several Mac Studios with unified memory and Thunderbolt 5 is becoming a working alternative for anyone who cares about private inference of large MoE models without the cloud. It can't yet be called a mass-market solution — software immaturity and setup complexity remain barriers.

Frequently Asked Questions

Why combine several Mac Studios into a cluster?

To get a single 1.5 TB shared memory pool. Models like DeepSeek V3 and Kimi K2 Thinking, at a trillion parameters, don't fit in the memory of one graphics card, while four Mac Studios with shared memory and RDMA over Thunderbolt 5 provide that much volume.

Which models were successfully run locally?

The author ran language models at the level of DeepSeek V3 and Kimi K2 Thinking on the cluster — both around 1 trillion parameters. The 1.5 TB shared-memory configuration was assembled specifically for them.