What RDMA over Thunderbolt 5 is, and which Macs have it
Apple’s technote TN3205 says it plainly: “RDMA over Thunderbolt is available starting with macOS 26.2 on Macs with Apple silicon with Thunderbolt 5.” It exposes an RDMA Verbs interface on the Thunderbolt controller, so a distributed runtime can move tensors between Macs at the hardware’s limit instead of through the TCP/IP stack. Ordinary IP over Thunderbolt keeps working beside it.
The Thunderbolt 5 requirement rules out a lot of Macs. From Apple’s spec pages:
| Mac | Thunderbolt 5 ports | RDMA? |
|---|---|---|
| Mac Studio M5 Ultra | 6 (four on the back, two on the front) | Yes |
| Mac Studio M5 Max | 4 (the front ports are USB-C) | Yes |
| Mac mini M5 Pro | 3 | Yes |
| MacBook Pro M5 Pro or M5 Max | 3 | Yes |
| Mac mini M6 | none (three Thunderbolt 4) | No |
| Used: Mac Studio M3 Ultra or M4 Max, Mac mini M4 Pro | Thunderbolt 5 | Yes |
| Used: Mac Studio M1 or M2 Ultra, M1–M3 Max MacBook Pro | Thunderbolt 4 | No |
Thunderbolt 4 Macs can still be clustered over plain TCP with MLX’s ring backend or a network, at much higher latency; this guide is about RDMA. Every Mac in the cluster should run the same macOS and the same runtime version.
Cabling: one cable per pair of Macs
RDMA over Thunderbolt is point-to-point: there is no Thunderbolt switch, and the link cannot route through a third Mac on its own. TN3205 recommends connecting every Mac to every other one (“fully-connected”) for the lowest latency, and MLX’s jaccl backend supports only fully connected topologies. So the cable count grows quickly:
| Macs | Cables (one per pair) | Ports used on each Mac | Fits |
|---|---|---|---|
| 2 | 1 | 1 | any Thunderbolt 5 Mac |
| 3 | 3 | 2 | any Thunderbolt 5 Mac |
| 4 | 6 | 3 | Mac mini M5 Pro (all three ports), Mac Studio |
| 4, doubled links | 12 | 6 | Mac Studio M5 Ultra or M3 Ultra (six ports each) |
MLX’s documentation notes that two Macs “can also be connected with more than one thunderbolt cable”, and Apple’s four-node cluster used two cables per pair. In its tests the second cable mattered for long prompts: with Muse Spark Open on four M5 Ultras at a 65,536-token prompt, time to first token fell from about 50.5 to 40.6 seconds. TN3205 also shows a five-Mac ring, which needs the application to forward data between Macs that are not directly linked. Use certified Thunderbolt 5 cables; a USB-C charging cable will not carry the link.
Turn it on: Recovery, rdma_ctl and the Thunderbolt Bridge
On each Mac, following TN3205:
- Update to macOS 26.2 or later.
- Start up in macOS Recovery, open Utilities → Terminal and run
rdma_ctl enable. MLX’s docs stress that this “cannot be done remotely even with sudo”. - Restart into macOS and check the interfaces:
ibv_deviceslists onerdma_endevice per Thunderbolt port, andibv_devinfoshowsPORT_ACTIVEonce the Mac at the other end has RDMA enabled too. - Turn the Thunderbolt Bridge off: System Settings → Network → Thunderbolt Bridge → Make Inactive. TN3205 requires it for loops (the bridge forwards frames between ports like a hub, and in a loop they circulate indefinitely); MLX’s jaccl requires it always, even for two Macs, plus a separate local network for each Thunderbolt link.
mlx.distributed_config --auto-setupdoes both. - Make sure every Mac can reach the others over SSH on Ethernet or Wi-Fi; the launch tools use it to start the processes.
Run a model across Macs with MLX (jaccl)
MLX’s RDMA backend is called jaccl and needs MLX 0.30.1 or later. MLX ships a helper that SSHes into each Mac, reads the Thunderbolt connections, checks that they form a valid mesh and writes the host file; then mlx.launch starts the same command on every Mac. From the MLX distributed docs and Apple’s post:
pip install -U mlx mlx-lm # on every Mac, same versions
# discover the Thunderbolt mesh and write a host file (needs passwordless sudo for --auto-setup)
mlx.distributed_config --verbose --hosts studio-1,studio-2 \
--over thunderbolt --backend jaccl --auto-setup --output cluster.json
# run one model across both Macs
mlx.launch --verbose --backend jaccl --hostfile cluster.json -- \
/path/to/python -m mlx_lm chat --model mlx-community/<model>
Two ways to split a model, and they do different things. Pipeline parallelism gives each Mac a slice of the layers: it is how a model too big for one Mac fits at all, but each token still passes through every slice in turn, so decode speed stays about where one Mac would be. Tensor parallelism splits every layer across the Macs, so they read their shares of the weights at the same time: decode gets faster, at the cost of more traffic over the cables. In mlx_lm.benchmark Apple switched pipeline mode on with --pipeline. Apple’s post also sets MLX_METAL_FAST_SYNCH=1; MLX’s docs now warn it can deadlock the GPU (#3142) and leave it off by default, so the command above omits it.
exo: the same idea with a GUI
exo wraps MLX’s distributed backends in an app that discovers the other Macs and places the model for you. Its README claims “day-0 support for RDMA over Thunderbolt 5” and says it needs macOS Tahoe 26.2 or later; it installs with brew install --cask exo or as a disk image, and offers pipeline and tensor sharding. Its README does not list M5 Macs among its RDMA hardware as of 18 November 2026, and its performance charts are credited to a third-party blog, so for the M5 Ultra the reference figures are still Apple’s.
What Apple measured: two and four M5 Ultras
Apple Machine Learning Research published the reference figures on 15 October 2026 in “Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio”, using MLX with jaccl over Thunderbolt 5. These are cluster results, a different workload from the single-Mac rows in the LLMCheck index, and are quoted here as Apple published them rather than entered as index rows.
Muse Spark Open (816B parameters, 42B active, run in NVFP4) needs about 493 GB for its weights alone and, Apple says, does not fit on one machine with its intermediate computations. Two M5 Ultras decoded it at 15.74 tok/s with pipeline parallelism and 24.99 with tensor parallelism; four M5 Ultras with tensor parallelism reached 36.73 tok/s, with the first token of the benchmark prompt after 4.8 seconds.
Mistral Medium 3.5 (128B dense, FP8) fits one M5 Ultra and shows the scaling cleanly: sourced7.70 tok/s on one Mac, 7.70 on two in pipeline mode, 13.69 on two in tensor mode and 23.37 on four, while the time to first token fell from 24.6 seconds on one Mac to 7.5 on four. Pipeline mode left decode unchanged, as expected, and Apple reports it up to about 1.8× faster than one Mac at prefilling a 65,536-token prompt. At that prompt length, two-Mac tensor parallelism was slower than pipeline for Muse Spark Open.
Apple also gives the power: about 300 W peak per M5 Ultra at a 65,536-token prompt (about 200 W for an M3 Ultra), so by LLMCheck’s arithmetic (four times Apple’s one-Mac figure) a four-Mac cluster draws around 1.2 kW at peak. For single-Mac context: in the same post an M5 Ultra decoded Qwen 3.8 27B (BF16) at sourced19.69 tok/s against sourced12.46 on an M3 Ultra, and the index’s figure for Qwen3.8-27B at 4-bit on an M5 Ultra is sourced48 tok/s.
Which models in the index need two Macs
The index’s minimum RAM is a model’s smallest build of at least 3 bits with a working Mac runtime, plus about 2 GB, except where the only published Mac run is at 2-bit. Anything above 512 GB needs more than one Mac; so does a model you want at 4-bit when its 4-bit build passes 512 GB. The arithmetic below is on paper: whether MLX or exo supports a model’s architecture, and whether anyone has published a cluster run of it, is a separate question to check before buying.
| Model | Index min RAM | One 512 GB Mac? | Cluster | Note |
|---|---|---|---|---|
| Muse Spark Open | 342 GB (3-bit) | Yes, at 3-bit | 2 × 512 GB in NVFP4 | Apple’s test model: about 493 GB of NVFP4 weights, which Apple says does not fit one Mac with its working memory |
| Mistral Large 4 | 383 GB (2-bit) | Yes, at 2–3-bit | 2 × 512 GB at 4-bit (about 600 GB) | what one 512 GB Mac runs |
| Inkling | 550 GB | No | 2 × 512 GB | Mac runtime caveat: llama.cpp support is an unmerged pull request |
| MiMo-V2.6-Pro | 581 GB | No | 2 × 512 GB | |
| DeepSeek V4 Pro | 850 GB | No | 2 × 512 GB | tight: 1,024 GB in total |
| Qwen3.8-2.4T-A95B | 1,310 GB | No | 3–4 × 512 GB | |
| Kimi K3 and Kimi K3.1 Pro | about 1,560 GB | No | 4 × 512 GB | 2.8T parameters; listed as server-class in the index |
Sizes from the LLMCheck catalog on 18 November 2026; Mistral Large 4 at 4-bit uses the index’s Q4_K_M sizing (0.57 GB per billion parameters, plus 2 GB). macOS lets the GPU use only part of each Mac’s memory by default, so a model within a few percent of the total needs iogpu.wired_limit_mb raised or one more Mac, and a long context adds its KV cache on top (methodology).
Everything the index ranks below 512 GB runs on one Mac, and one Mac is simpler and usually faster per dollar: the index’s pick from 192 to 512 GB, MiMo-V2.6-Flash, runs at community77 tok/s on a single M5 Ultra.
What a cluster costs, and when to rent instead
At Apple’s US list prices (checked 14 November 2026), two 512 GB M5 Ultra Mac Studios cost $44,598 and four cost $89,196, before cables. A cheaper way to 512 GB in total is two 256 GB M5 Ultras with the 64-core GPU at $9,499 each ($18,998), less than one 512 GB Mac ($22,299), with two Macs’ bandwidth for tensor parallelism; the catch is that each Mac then holds only half the model, so the split must fit 256 GB per Mac. The 512 GB M5 Ultra and every 256 GB configuration are build-to-order from Apple; AppleInsider’s price guide showed no Amazon listing for any M5 Ultra configuration on 18 November 2026.
Cables for a Thunderbolt 5 cluster
One certified Thunderbolt 5 cable per pair of Macs: 1 for two Macs, 3 for three, 6 for four (12 to double every link on M5 or M3 Ultras). Short cables keep the desk tidy; check the listing says Thunderbolt 5, not only USB-C.
Thunderbolt 5 cables on Amazon →Disclosure: As an Amazon Associate, LLMCheck earns from qualifying purchases. This doesn’t affect the index’s figures or rankings.
Rent instead: a GPU server by the hour
For a model you want to try rather than run every day, or one no Mac cluster holds at full precision, renting a multi-GPU server costs far less than four Mac Studios. Vast.ai rents datacenter GPUs by the hour; compare its prices with other providers before you commit.
Referral link: if you sign up through it, LLMCheck earns a share of what you spend at Vast.ai, at no extra cost to you. It never affects the index’s figures or rankings.
Related guides
- What a 512 GB Mac Studio runs, on one Mac.
- Best LLMs for the Mac Studio M5 Ultra 512 GB.
- Apple Silicon chips compared, including Apple’s M5 Ultra vs M3 Ultra figures.
- Best used Mac for local LLMs, and which used Macs have Thunderbolt 5.