Which Mac Studio Should You Buy for Local LLMs? (2026)

Bottom Line

• Cheapest config that runs gpt-oss 120B: M5 Max, 40-core GPU, 128GB, 512GB SSD, $5,099.
• Fastest config under $6,000: M5 Ultra, 64-core GPU, 96GB, 1TB SSD, $5,499. Same model as the 128GB Max, about 2x the token speed.
• The 256GB tier finally has a reason: M5 Ultra, 64-core GPU, 256GB, $9,499. It is the first Apple tier that loads MiMo-V2.6-Flash, which peaks at about 165GB in community runs.
• Never order the 32-core GPU M5 Max. It is sold with 36GB only and runs memory at 460 GB/s. Every 48GB, 64GB and 128GB Max uses the 40-core GPU at 614 GB/s.
• Skip the $1,300 Ultra GPU step. Apple lists 1.2 TB/s for both the 64-core and 80-core GPU. That step buys prompt processing, not token speed.
• Skip Apple storage. Every SSD step costs more per GB than an external drive. A 1TB step on the Max is $300.

Prices are Apple US prices, as of September 2026, read from Apple’s US configurator on 21 September 2026. Machines ship starting 22 September 2026. The M4 Max and M3 Ultra are discontinued.

Ready to buy? See the tested hardware list with current prices.

The Decision Nobody Frames Correctly

Most Mac Studio buying pages compare Max and Ultra as two products. Apple’s configurator sells eight configurations, and only two numbers separate them for local AI. Unified memory sets what fits. Memory bandwidth sets how fast a model reads its weights each token. CPU cores and GPU cores matter far less than either.

Bandwidth comes in three fixed steps: 460 GB/s, 614 GB/s and 1.2 TB/s. Memory comes in six sizes: 36GB, 48GB, 64GB, 96GB, 128GB and 256GB. The 512GB M5 Ultra is listed as “coming late October” with no price, and cannot be pre-ordered.

The model picks for each tier are on the Mac Studio local LLM hub. This page only tells you which line to click.

Every Configuration, Priced

All prices are US, base storage, as of September 2026. Price per GB is a derived figure: price divided by unified memory in GB. Usable memory assumes macOS gives the GPU roughly 75% of unified memory, an approximation.

Max configs ship with 512GB of storage. Ultra configs ship with 1TB. The “Runs” column comes from the hub page’s tier table: 36GB to 64GB is the 27B class, 96GB and 128GB run gpt-oss 120B at about 65GB of weights, and 256GB runs 235B-class MoE.

The arithmetic for the derived column: $2,499 ÷ 36 = $69.4, $3,099 ÷ 48 = $64.6, $3,499 ÷ 64 = $54.7, $5,099 ÷ 128 = $39.8, $5,499 ÷ 96 = $57.3, $6,799 ÷ 96 = $70.8, $9,499 ÷ 256 = $37.1, $10,799 ÷ 256 = $42.2.

Two things stand out. The 256GB Ultra with the 64-core GPU is the cheapest memory in the whole line per GB. The 36GB base is the most expensive, and it also has the slowest bus.

The Three Traps in the Configurator

Trap 1: the 32-core GPU

The $2,499 base Mac Studio uses the 32-core GPU M5 Max. Apple lists it at 460 GB/s. Every other M5 Max config uses the 40-core GPU at 614 GB/s. The configurator has no 64GB or 128GB page for the 32-core GPU; those URLs redirect. So the 48GB step is also a bandwidth step. The $600 from $2,499 to $3,099 buys 12GB of memory and a bus that is a third faster (derived: 614 ÷ 460 = 1.33). Do not buy the 36GB base for local AI.

Trap 2: the 80-core GPU Ultra

The step from the 64-core to the 80-core GPU costs $1,300 at both 96GB and 256GB (derived: $6,799 minus $5,499, and $10,799 minus $9,499). Apple’s specs page lists 1.2 TB/s for both. Token generation is bandwidth-bound, so the 80-core GPU generates tokens at the same speed as the 64-core. The extra cores speed up prompt processing, which is compute-bound. If you feed long documents to a model all day, it may pay. If you chat and run agents, it buys nothing you can measure per token. Put the $1,300 toward memory instead.

Trap 3: Apple storage

Apple’s storage steps, as reported on 25 August 2026: M5 Max 1TB is +$300, 2TB is +$800, 4TB is +$1,800, 8TB is +$3,800. M5 Ultra 2TB is +$500, 4TB is +$1,500, 8TB is +$3,500, 16TB is +$7,500. The fully loaded machine is $18,299. Model weights load into memory once and then stream from nowhere. A Thunderbolt NVMe enclosure holds the model library at a fraction of these prices. Buy the base storage and spend on memory.

One Pick Per Budget

1. Under $3,500: M5 Max, 40-core GPU, 64GB, $3,499. A 27B-class machine at 614 GB/s. Skip the 36GB and 48GB steps; 64GB is the only Max under $5,000 with room for a 27B at Q8 plus context. gpt-oss 120B does not fit.
2. Around $5,000, capacity first: M5 Max, 40-core GPU, 128GB, $5,099. The cheapest new Apple box that runs gpt-oss 120B at full 128K context with a second model resident. About $39.8 per GB, derived.
3. Around $5,500, speed first: M5 Ultra, 64-core GPU, 96GB, $5,499. Runs the same gpt-oss 120B at about 2x the token speed of the Max (derived: 1.2 TB/s ÷ 614 GB/s = 1.95). You give up 32GB of headroom for $400 more. If the model you want is under 65GB, this is the better $5,000-class box.
4. Around $9,500: M5 Ultra, 64-core GPU, 256GB, $9,499. The only config on this page with a new, concrete reason to exist. See the next section.
5. The 80-core GPU at any memory size: no. Same 1.2 TB/s. Save the $1,300.
6. 512GB: wait. Late October, no price, no pre-order. Our wait-for-M5-Ultra guide covers the timing.

Why 256GB Now Has a Concrete Reason

Until this week the honest advice was that the 96GB to 256GB step bought capacity nobody had a named model for. That changed on 21 September 2026.

Xiaomi released MiMo-V2.6-Flash-RL that day on Hugging Face (XiaomiMiMo/MiMo-V2.6-Flash-RL, MIT license). It is a 309B-total, 15B-active mixture-of-experts model with a 1M-token context. The native checkpoint is 165.53 GiB with MXFP4 experts and FP8 attention.

The community-reported Mac numbers, from the Hugging Face model pages named here:

• The MLX 4-bit conversion Vontra/MiMo-V2.6-Flash-RL-MLX-4bit-MTP, measured on a 256GB M3 Ultra Mac Studio: 59.4 tok/s sustained decode, 477 to 563 tok/s prompt processing, peak unified memory 164.3GB on a short prompt and 166.8GB on a 2,048-token prompt.
• mlx-community/MiMo-V2.6-Flash-RL-mxfp4-q8 is 156GB on disk, and its card says loading takes about 170GB of unified memory.

A 128GB Mac Studio can allocate roughly 96GB to the GPU under the 75% approximation. A 165GB model does not fit, at any setting. The 256GB tier, at roughly 192GB usable, holds it with about 25GB to spare (derived: 192 minus 167). So 256GB is the first Apple tier that runs a 15B-active model that decodes at nearly 60 tok/s on the older M3 Ultra. Those figures are community-reported on an M3 Ultra, not an M5 Ultra; the M5 Ultra’s 1.2 TB/s should raise decode speed, but no one has measured it yet.

The full fit check, including the 128GB and PC options, is in Can I run MiMo-V2.6-Flash locally?.

The Discontinued Option

Apple stopped selling the M4 Max Mac Studio on 25 August 2026. It has 546 GB/s of bandwidth, which is 89% of the M5 Max’s 614 GB/s, and its 128GB tier runs gpt-oss 120B today. If a retailer listing still shows stock and the listing says 128GB, the Mac Studio M4 Max 128GB is worth checking against the $5,099 M5 Max 128GB. The 36GB and 64GB M4 Max units share the same name, so read the memory line before you order. The hub page’s honest recommendation covers this trade in more detail.

How to Read Your Own Order Page

Before you click Buy, check three lines on Apple’s configurator:

1. GPU cores on the Max. If it says 32-core, the memory is 36GB and the bus is 460 GB/s. You want 40-core.
2. GPU cores on the Ultra. If it says 80-core, you paid $1,300 for prompt processing. 64-core has the same 1.2 TB/s.
3. Storage. If it is above the base tier, you paid Apple’s step prices for space an external drive provides. Model weights do not need the internal SSD.

Memory is soldered. Order the tier for the largest model you intend to run in three years, not the one you run today.

See Also

• Best Local LLM for Mac Studio — the hub: model picks for every memory tier
• Can I Run MiMo-V2.6-Flash Locally? — the 165GB model that gives 256GB its reason
• Best Local LLM for Mac Studio M5 Ultra 256GB — the top tier in depth
• Best Models for the 256GB Mac Studio — the 235B-class table
• Mac mini vs Mac Studio for Local LLMs — whether you need the Studio at all
• Should You Wait for the M5 Ultra Mac Studio? — the 512GB timing question
• Mac Studio vs RTX Workstation for Local LLMs — Apple bandwidth vs CUDA