← All guides

Claris Harbor: Which Mac to Buy (2026): Memory First

Claris, an Apple company, has a private preview of Harbor, an app that turns your Mac into a private AI server and pools memory across Macs over Thunderbolt 5. Nobody outside the preview has a spec sheet yet. This page sorts what is confirmed from what is not, and tells you which Mac to buy if you want Harbor, or anything like it, to run well.

Bottom Line

Buy memory first. Buy Thunderbolt 5 only if you want to add a second Mac later. Harbor is a Claris app in private preview. Claris says it turns “the Apple silicon computers you own into a private AI server” and can “connect multiple machines over Thunderbolt 5 to pool unified memory” (Claris). There is no public download, no spec sheet and no price.

If you wantBuyWhy
One Mac as a private AI serverThe most unified memory you can afford, any current Apple silicon MacMemory decides which model fits. Claris lists no minimum chip
A Mac you can cluster laterMac mini M5 Pro, MacBook Pro M5 Pro or M5 Max, Mac Studio M5 Max or M5 UltraHarbor clusters over Thunderbolt 5 only
To avoid a dead endSkip the M6 Mac mini and M6 MacBook Pro for clusteringBoth have Thunderbolt 4 (apple.com specs)
A 120B-class model on one box96GB or moregpt-oss 120B in MLX format is 62.33 GB before context

Choosing a Mac for local AI? See the tested hardware list with current prices by memory tier.

What Harbor Is, and What Is Confirmed

Two reports broke the story on 10 November 2026. 9to5Mac quoted the listing: “Connect two or more Thunderbolt 5-capable Mac computers and Harbor pools their unified memory into a single inference resource.” MacRumors added that Harbor “is in early access and is not publicly available”, and that an App Store listing appeared and was removed.

QuestionStatus (11 November 2026)Source
Who makes itClaris, an Apple companyClaris page
Single Mac or clusterBoth: “an endpoint in minutes on a single Mac”, clusters over Thunderbolt 5Claris page
Gives you an API endpointYes, a “Copy API” button and a local URLLeaked App Store screenshots
API format (OpenAI-compatible?)Not statednone
Model formatNot stated. The screenshots load mlx-community/gpt-oss-20b-MXFP4-Q4 from Hugging Face, an MLX repoAaron Perris on X, 10 Nov
Minimum chip, RAM, macOSNot statednone
PriceNot stated. The early-access form asks for a company nameMacRumors
Release dateNot statednone

The screenshots come from Aaron Perris on X, who posted them from the pulled listing. They are marketing images, not measurements. iDropNews reports the app was “formerly known as Fabric”. The screenshot’s download folder, com.claris.fabric, matches that.

Which Macs Have Thunderbolt 5

This is the one hard requirement anyone has published. Apple’s current tech specs pages, read 11 November 2026:

MacThunderboltMax unified memoryMemory bandwidthCan join a Harbor cluster
Mac mini M6Thunderbolt 432GB153 GB/s (16GB), 170 GB/s (24/32GB)No
Mac mini M5 ProThunderbolt 5 (3 ports)64GB307 GB/sYes
MacBook Pro M6Thunderbolt 432GBup to 170 GB/sNo
MacBook Pro M5 ProThunderbolt 5 (3 ports)64GB (20-core GPU)307 GB/sYes
MacBook Pro M5 MaxThunderbolt 5 (3 ports)128GB (40-core GPU)460 or 614 GB/sYes
Mac Studio M5 MaxThunderbolt 5 (4 rear)128GB460 or 614 GB/sYes
Mac Studio M5 UltraThunderbolt 5 (4 rear, 2 front)512GB1.2 TB/sYes

Older M4 Pro and M4 Max Macs also have Thunderbolt 5. Apple’s own RDMA technote uses “a cluster of four M4 Pro Mac minis” as its example, and Harbor’s cluster screenshot shows two M4 Pro Mac minis. Apple no longer sells them new.

The trap is the M6. It is the cheapest current Mac mini and MacBook Pro, and it cannot join a cluster at all. If clustering is part of your plan, the entry point is the M5 Pro Mac mini, from $1,699 with 24GB (apple.com, as of November 2026).

RDMA: The Setup Step Nobody Mentions

MacRumors ties Harbor’s clustering to RDMA over Thunderbolt 5, which Apple added in macOS Tahoe 26.2. Apple’s technote TN3205 says three things a buyer should know:

  1. “RDMA over Thunderbolt is available starting with macOS 26.2 on Macs with Apple silicon with Thunderbolt 5.”
  2. You turn it on from macOS Recovery, by running rdma_ctl enable. It is off by default.
  3. “Disable the Thunderbolt bridge when connecting Macs in loops.”

What the screenshot actually shows. Harbor’s cluster image draws an M5 Max MacBook Pro (48GB) wired to two M4 Pro Mac minis (16GB and 24GB) over “TB5” links. The side panel in the same image reads Mode “Standalone”, RDMA “Not Enabled”, and Cluster Memory “48 GB”. That is the laptop’s own memory, not the 88GB the three machines hold together. Read it as a mockup, not proof that pooling works today.

What Fits in Each Memory Tier

Harbor has not published a model list. The screenshots show MLX-format files, so these sizes are MLX weights from the Hugging Face API, read 11 November 2026. Context and macOS need memory on top of the weights.

Model (MLX)WeightsSmallest Mac that makes sense
gpt-oss 20B, mlx-community/gpt-oss-20b-MXFP4-Q4 (the one in Harbor’s screenshot)11.18 GB24GB. It loads on 16GB, with little room left
Qwen 3.8 27B, mlx-community/Qwen3.8-27B-4bit16.05 GB32GB
Qwen 3.6 35B-A3B, mlx-community/Qwen3.6-35B-A3B-4bit20.40 GB32GB, 48GB for long context
gpt-oss 120B, mlx-community/gpt-oss-120b-MXFP4-Q462.33 GB96GB or 128GB

Pooling is the reason Harbor exists: a 120B model on two 64GB Macs instead of one 128GB Mac. No one has published Harbor cluster speeds yet. A Thunderbolt 5 port tops out at 120Gb/s, about 15 GB/s, while the M5 Pro’s memory runs at 307 GB/s (derived from apple.com specs). A model split across Macs pays that gap on every token it passes between them. Until someone measures it, one Mac with enough memory is the safe buy.

THE TIER THAT RUNS 27B-35B MODELS WITH CONTEXT

48GB of unified memory holds Qwen 3.8 27B or Qwen 3.6 35B-A3B with room for a long context. Check that the config you pick has Thunderbolt 5 if you plan to cluster.

MacBook Pro M5 Pro 48GB ↗

One Big Mac or Two Small Ones?

PlanWhat you getRisk
One Mac with 96-128GBRuns 120B-class models today, in any app: LM Studio, Ollama, llama.cpp, MLXHighest single price
Two smaller Thunderbolt 5 MacsPooled memory, if Harbor ships as describedNo published speeds; RDMA setup in Recovery; Harbor may not be free
An M6 Mac now, cluster laterNothing. M6 has Thunderbolt 4Dead end for clustering

Harbor is not the only way to pool Macs. MacRumors compares it with exo and LM Studio. If you already own two Thunderbolt 5 Macs, Harbor costs you nothing to try once it opens. If you are buying new, buy one Mac with the memory your target model needs. See Mac mini vs Mac Studio for local LLMs for that choice.

What Is Still Unknown

  • Minimum chip, memory and macOS version. A leaked dark-mode screenshot shows Harbor on an “M2 Pro · 16GB” MacBook Pro, which suggests single-Mac mode does not need Thunderbolt 5. That is our reading of an image, not a stated requirement.
  • Whether the API is OpenAI-compatible. Coding agents and OpenClaw need that to connect.
  • Price, release date, and whether cluster mode needs RDMA turned on.
  • Speed. There are no Harbor benchmarks. Commenters in the Hacker News thread (522 points) gave none either. One reported running Qwen 3.8 27B in MLX on a 128GB Mac until it “hits the heat limit” on long jobs. That is a community claim.

This page will be updated when Claris publishes requirements.

Sources

See Also

Before you buy a Mac for local models, check the tested hardware list for current prices by tier.

Read next

Which Mac Studio Should You Buy for Local LLMs? (2026)
The exact Mac Studio configuration to order for local LLMs, as of September 2026. Eight M5 Max and M5 Ultra configs from $2,499 to $10,799, priced per GB of unified memory, with one pick per budget. The 256GB tier now has a concrete reason: MiMo-V2.6-Flash needs about 165GB and does not fit a 128GB Mac.
Mac mini vs Mac Studio for Local LLMs (2026): Which to Buy
Mac mini vs Mac Studio for local LLMs and OpenClaw: unified memory ceilings, bandwidth, what each runs, and a clear buying recommendation.
Mac Studio vs RTX Workstation for Local LLMs (2026): Which Should You Buy?
Mac Studio vs RTX workstation for local LLMs, Ollama, and OpenClaw: unified memory vs CUDA VRAM, speed, model fit, setup complexity, and buying recommendations.
Which Mac mini Should You Buy for Local LLMs? (2026)
The exact Mac mini configuration to order for local LLMs, as of September 2026. Nine M6 and M5 Pro configs from $899 to $2,899, priced per GB. Buy the 32GB M6 at $1,299. The $899 16GB base runs memory at 153 GB/s, not 170, and no Mac mini runs gpt-oss 120B.