Compare

Mac Studio for local AI

The Mac Studio is Apple’s small desktop with up to 512 GB of unified memory shared by the CPU and GPU, enough to hold large language models whole. This guide covers the M5 Max and M5 Ultra on sale now and the 2025 M4 Max and M3 Ultra people still look for, on the models a Lucebox runs.

M5 Max: $5,399 for 128 GB

The 40-core M5 Max reads memory at 614 GB/s and holds models of around 100 GB, such as DeepSeek V4 Flash (101.5 GB).

M5 Ultra: up to 512 GB

About 1.2 TB/s per Apple, and the only Mac Studio on sale that holds DeepSeek V4.1 Flash, from $10,799 with 256 GB.

M4 Max and M3 Ultra: not sold new

Apple’s store dropped both. The M3 Ultra still has the faster memory of the two, 819 GB/s against 546.

Which Mac Studio chips are there?

The Mac Studio comes with four chips that matter for local AI: the M5 Max and M5 Ultra Apple sells today, and the 2025 M4 Max and M3 Ultra it sold before them.

Mac Studio (M5 Max)Mac Studio (M5 Ultra)Mac Studio (M4 Max, 2025)Mac Studio (M3 Ultra, 2025)Lucebox Zero 495
Chip Apple M5 Max, 18 CPU cores, 32 or 40 GPU coresApple M5 Ultra, 30 or 36 CPU cores, 64 or 80 GPU coresApple M4 Max, 14 or 16 CPU cores, 32 or 40 GPU coresApple M3 Ultra, 60 or 80 GPU coresAMD Ryzen AI MAX+ PRO 495 and a Radeon AI PRO R9700 card
Memory 36 GB on the 32-core GPU; 48, 64 or 128 GB on the 40-core96 GB; 256 or 512 GB with the 80-core GPU36 GB on the 32-core GPU; 48, 64 or 128 GB on the 40-core96 to 512 GB at launch32 GB on the card, plus 128 GB or 192 GB LPDDR5X-8533
Memory speed 614 GB/s (460 on the 32-core GPU)1.2 TB/s546 GB/s (410 on the 32-core GPU)819 GB/s640 GB/s on the card, 273 GB/s unified
Networking 10 GbE, Wi-Fi10 GbE, Wi-Fi10 GbE, Wi-Fi 6E10 GbE, Wi-Fi 6E10 GbE, Wi-Fi 7, optional 100 GbE cluster kit
Ports and expansion 4x Thunderbolt 5, USB-A, HDMI 2.14x Thunderbolt 5, USB-A, HDMI 2.14x Thunderbolt 5, 2x USB-C, 2x USB-A, HDMI 2.1, SD card slot6x Thunderbolt 5, 2x USB-A, HDMI 2.1, SD card slot2× USB4, 2× USB-A 3.2 Gen 2, 2× USB-A 2.0, HDMI 2.1, DisplayPort 1.4, plus 4 display outputs on the GPU
Power Up to 480 WUp to 480 WUp to 480 WUp to 480 W1,000 W power supply, SFX form factor
Size and weight 95 x 197 x 197 mm, 2.7 kg95 x 197 x 197 mm, 3.6 kg95 x 197 x 197 mm, 2.74 kg95 x 197 x 197 mm, 3.64 kg340 × 110 × 320 mm, 11.97 L, 6 kg
Ships with macOSmacOSmacOSmacOSUbuntu Server with the Lucebox engine and models installed
Warranty Not statedNot statedNot statedNot stated1 year, parts and labor
Sources Apple Apple Apple (2025 model) Apple (2025 model), Apple, March 2025 Lucebox

Mac Studio M5 Ultra

The Mac Studio M5 Ultra has 30 or 36 CPU cores, 64 or 80 GPU cores and, per Apple, about 1.2 TB/s of memory bandwidth. It comes with 96 GB, or 256 or 512 GB with the 80-core GPU, from $5,499 for 96 GB to $22,299 for 512 GB on Apple’s US store (31 October 2026).

Mac Studio M5 Max

The Mac Studio M5 Max has an 18-core CPU and a 32 or 40-core GPU; the 40-core chip reads memory at 614 GB/s and takes 48, 64 or 128 GB. With 128 GB and a 1 TB drive it costs $5,399, and it is the Mac in our matched DGX Spark vs Mac Studio run.

M3 Ultra vs M4 Max: which Mac Studio for AI?

For local AI the Mac Studio M3 Ultra is the stronger pick: it reads memory at 819 GB/s, 1.5x the 546 GB/s of a 40-core M4 Max, and it launched with up to 512 GB where the M4 Max stops at 128 GB.

The M4 Max is enough for models of around 100 GB: Qwen3.8-27B (15.2 GB), DeepSeek V4 Flash (101.5 GB) and Qwen3.8-Flash-Next (106.3 GB) each fit in 128 GB. The M3 Ultra adds DeepSeek V4.1 Flash (about 191 GB) at 256 GB, and room for two large models at once at 512 GB.

We have no M4 Max run in our data, so the gap in writing speed is an estimate from bandwidth. Apple sells neither model new any more.

Which models fit: M4 Max 128 GB, M3 Ultra 256 GB and 512 GB

A 128 GB Mac Studio M4 Max holds DeepSeek V4 Flash (101.5 GB) and Qwen3.8-Flash-Next (106.3 GB); DeepSeek V4.1 Flash (about 191 GB) needs a 256 or 512 GB M3 Ultra or M5 Ultra.

Model and size Mac Studio (M4 Max, 2025)128 GB unifiedMac Studio (M3 Ultra, 2025), 256 GB256 GB unifiedMac Studio (M3 Ultra, 2025), 512 GB512 GB unifiedMac Studio (M5 Max), 128 GB128 GB unifiedMac Studio (M5 Ultra), 256 GB256 GB unifiedLucebox Zero 495, 192 GB32 GB card + 192 GB
Qwen3.8-27B15.2 GB YesYesYesYesYesYes, on the card
DeepSeek V4 Flash101.5 GB YesYesYesYesYesYes, card and unified
Qwen3.8-Flash-Next106.3 GB YesYesYesYesYesYes, card and unified
Qwen3.8-27B and DeepSeek V4 Flash at once116.7 GB Just, little room leftYesYesJust, little room leftYesYes, card and unified
DeepSeek V4.1 Flash190.8 GB NoYesYesNoYesYes, card and unified

How fast is a Mac Studio for local LLMs?

A Mac Studio M5 Max writes DeepSeek V4 Flash at 39.35 tokens per second on a 2K prompt in antirez’s ds4 sweep, and one LocalMaxxing user reports about 33 tokens per second on a 512 GB M3 Ultra. The chart sets the Mac Studio runs beside a DGX Spark, a Strix Halo box and the Lucebox.

Model
Measure
  1. M5 Max, 2K prompt 39.35 tok/s DwarfStar (ds4) · Q2 (ds4) · 2,048-token prompt · no drafting Public
  2. M5 Max, 16K prompt 36.14 tok/s DwarfStar (ds4) · Q2 (ds4) · 16,384-token prompt · no drafting Public
  3. M5 Max, 32K prompt 34.36 tok/s DwarfStar (ds4) · Q2 (ds4) · 32,768-token prompt · no drafting Public
  4. M3 Ultra 33 tok/s ds4 on Metal · IQ2_XXS · no drafting LocalMaxxing
  5. Strix Halo 30.75 tok/s hipfire · MQ2R (2.74 bits) · 128-token prompt · no drafting LocalMaxxing
  6. M5 Max, 64K prompt 27.64 tok/s DwarfStar (ds4) · Q2 (ds4) · 65,536-token prompt · no drafting Public
  7. DGX Spark 18.05 tok/s DwarfStar (ds4) · Q2 (ds4) · 2,048-token prompt · no drafting Public

M5 Max writes 2.2x as fast as DGX Spark with a 2,048-token prompt, same engine and file. The other runs used a different setup.

Mac Studio speed test: watch them write

Race
Task

Each machine's own published setup, no drafting. Engines, files and prompt lengths differ.

Mac Studio M5 Max0.0 s
 
writing at 39.35 tok/s
 
 
Mac Studio M3 Ultra0.0 s
 
 
writing at 33 tok/s
 
 
Strix Halo box0.0 s
 
 
writing at 30.75 tok/s
 
 
DGX Spark0.0 s
 
 
writing at 18.05 tok/s
 
 
 
  
  
 
 
 
 
Mac Studio M3 Ultra and M4 Max price
 
Apple no longer sells the Mac Studio M3 Ultra or M4 Max new, so neither has an Apple price in November 2026; the M5 Max that replaced the M4 Max costs $5,399 with 128 GB. Before the M3 Ultra left the store, Apple raised its 96 to 256 GB upgrade from $1,600 to $2,000 in March 2026, press reports say.
 
 
  
 
  
 Machine Memory Bandwidth, GB/s Software US price 
   
NVIDIA GB10
 
  NVIDIA DGX Spark   128 GB 273 CUDA      $6,950 to $6,995      
  
Apple Mac Studio
 
  Mac Studio (M5 Max)   48 or 128 GB 614 MLX, Metal    48 GB:   $3,099    128 GB:   $5,399      
  Mac Studio (M5 Ultra)   96 or 256 or 512 GB 1,228 MLX, Metal    96 GB:   $5,499    256 GB:   $10,799    512 GB:   $22,299      
  Mac Studio (M4 Max, 2025)   128 GB 546 MLX, Metal  Apple no longer sells it new    
  Mac Studio (M3 Ultra, 2025)   256 or 512 GB 819 MLX, Metal  Apple no longer sells it new    
  
Lucebox
 
  Lucebox Zero 495, 128 GB   32 GB on the GPU + 128 GB 640 / 273 ROCm, Vulkan      $5,999      
  Lucebox Zero 495, 192 GB   32 GB on the GPU + 192 GB 640 / 273 ROCm, Vulkan      $8,199      
  
 
US prices read on 30 and 31 October 2026.
 
   
  
 
 
 
Other machines for the price of the Mac Studio
  
 
 
 
 
 
 Model 
  
 
 
  
  $5,999 
 
 
 Rank by 
   
 
 
 
 
 
 
Our pick
 
Lucebox Zero 495, 128 GB
 
  • Holds DeepSeek V4 Flash (101.5 GB) across the 32 GB card and 128 GB of unified memory.
  • Writes it at 51.1 tok/s on our earlier 395 build, drafting on (our lab).
  • $5,999, right at your budget.
 
 
Also consider
 
GMKtec EVO-X2, 128 GB
 
  • Holds DeepSeek V4 Flash (101.5 GB) in 128 GB of unified memory.
  • Writes it at 41.9 tok/s on another box with the same chip, drafting on (our lab).
  • $3,500, $2,499 under your budget.
 
 
We make the Lucebox. It is ranked by the same rule as every other row; a run marked “on our earlier 395 build” comes from the previous Lucebox, with the same R9700 card.
 
 
  
 
 
  
     
  1.  
    Lucebox Zero 495, 128 GB32 GB card + 128 GB
     
      Card + unified 63% used 
     
      $5,999 
     
      273 GB/s 
     
     51.1 tok/s drafting on, our lab, on our earlier 395 build 
     
  2.  
    GMKtec EVO-X2, 128 GB128 GB unified
     
      Fits 79% used 
     
      $3,500 
     
      256 GB/s 
     
     41.9 tok/s drafting on, our lab, on another box with the same chip 
     
  3.  
    Mac Studio (M5 Max), 128 GB128 GB unified
     
      Fits 79% used 
     
      $5,399 
     
      614 GB/s 
     
     39.35 tok/s no drafting, public benchmark 
     
  4.  
    GMKtec EVO-X2, 64 GB64 GB unified
     
      Does not fit needs 159% 
     
      $2,200 
     
      256 GB/s 
     
     no run yet  
     
  5.  
    Mac Studio (M5 Max), 48 GB48 GB unified
     
      Does not fit needs 211% 
     
      $3,099 
     
      614 GB/s 
     
     no run yet  
     
  6.  
    GB10 box with 64 GB (Dell, ASUS and others)64 GB unified
     
      Does not fit needs 159% 
     
      $4,999 
     
      273 GB/s 
     
     no run yet  
     
  7.  
    Mac Studio (M5 Ultra), 96 GB96 GB unified
     
      Does not fit needs 106% 
     
      $5,499 
     
      1,228 GB/s 
     
     no run yet  
     
  8.  
    Framework Desktop (Ryzen AI MAX+ PRO 495)192 GB unified$800 over budget
     
      Fits 53% used 
     
      $6,799 
     
      273 GB/s 
     
     no run yet  
     
  9.  
    NVIDIA DGX Spark128 GB unified$951 over budget
     
      Fits 79% used 
     
      $6,950 
     
      273 GB/s 
     
     35.3 tok/s drafting on, public benchmark 
     
  10.  
    RTX 5090 card32 GB on the card
     
      Does not fit needs 317% 
     
      $7,681 
     
      1,792 GB/s 
     
     no run yet  
     
  11.  
    Lucebox Zero 495, 192 GB32 GB card + 192 GB$2,200 over budget
     
      Card + unified 45% used 
     
      $8,199 
     
      273 GB/s 
     
     51.1 tok/s drafting on, our lab, on our earlier 395 build 
     
  12.  
    Mac Studio (M5 Ultra), 256 GB256 GB unified$4,800 over budget
     
      Fits 40% used 
     
      $10,799 
     
      1,228 GB/s 
     
     no run yet  
     
  13.  
    RTX PRO 6000 workstation96 GB on the card
     
      Does not fit needs 106% 
     
      $21,460 
     
      1,792 GB/s 
     
     no run yet  
     
  14.  
    tinybox red flash192 GB on 6 cards$16,001 over budget
     
      On the card 53% used 
     
      $22,000 
     
      3,840 GB/s 
     
     no run yet  
     
  15.  
    Mac Studio (M5 Ultra), 512 GB512 GB unified$16,300 over budget
     
      Fits 20% used 
     
      $22,299 
     
      1,228 GB/s 
     
     no run yet  
     
    16.  
 
 
The pick holds the model with room for the prompt, fits your budget and has the fastest run we can link to. Runs with drafting rank against each other first, then runs without it, then machines with no run, by memory speed. The amber band is the Lucebox's 32 GB card.
  
  
  
  
 
 
 
Who should buy a Mac Studio for AI?
  
 
  
Buy one if
  • You mostly chat and draft: writing speed follows memory bandwidth, and an M5 Max has 614 GB/s.
  • You want a quiet, small desktop and already work in macOS.
  • You need 256 or 512 GB in one machine: the M5 Ultra is the only Mac Studio on sale with that much ($10,799 for 256 GB).
Look elsewhere if
  • You paste long prompts all day: an M5 Max read a 64K-token prompt at 398.5 tok/s in the ds4 sweep, and prompt reading is where Macs are slow.
  • You run a 27B model most of the day: an R9700 card wrote Qwen3.8-27B at 303.38 tok/s with drafting on a short code prompt (36.7 without), against 56.4 on an M5 Max, each with drafting on its own engine.
  • Your software needs CUDA, which macOS does not run.
   
  
 
 
 
Mac Studio or Lucebox: where each fits
  
 
  
 
We make the Lucebox Zero 495, so read this box as ours.
 
 
 
Where it fits
 
  • DeepSeek V4.1 Flash with only 128 GB, reading rarely used experts from the SSD (26 tok/s at best while writing code in our lab), where a Mac Studio needs 256 GB to hold the model.
  • DeepSeek V4.1 Flash for $8,199 with 192 GB, where the cheapest Mac Studio that holds it is the $10,799 M5 Ultra with 256 GB.
  • Qwen3.8-27B on the card and DeepSeek V4 Flash in the unified memory, both loaded and answering together.
 
 
 
Where it doesn't
 
  • Silence and size come first: a Mac Studio is smaller and quieter.
  • You want more than 192 GB in one box: the M5 Ultra goes to 512 GB.
  • You need a machine before January.
 
 
 

From $5,999, delivered from January 2027.
See the hardware Configure and order 
 
   
   
 
 
 
More local AI hardware comparisons
 
 
 
 
 
NVIDIA DGX Spark
  
 
AMD Ryzen AI MAX+ boxes
  
 
Apple Mac Studio
  
 
Graphics card builds
  
 
 
  
 
 
 
Mac Studio FAQ
 
 
 
 
 
Is the Mac Studio M4 Max good for local AI?
 
Yes, for models of around 100 GB: the 2025 Mac Studio M4 Max reads memory at 546 GB/s with the 40-core GPU and holds up to 128 GB, enough for DeepSeek V4 Flash (101.5 GB). It runs the same models as the M5 Max at 614 GB/s, a little slower, and Apple no longer sells it new.
 
 
What is the M3 Ultra memory bandwidth?
 
The Apple M3 Ultra in the 2025 Mac Studio has 819 GB/s of memory bandwidth, per Apple’s spec page. That is more than the M5 Max’s 614 GB/s and less than the M5 Ultra’s, which Apple lists at about 1.2 TB/s.
 
 
Is the Mac Studio M3 Ultra good for running LLMs?
 
Yes, for large models: the Mac Studio M3 Ultra reads memory at 819 GB/s and launched with up to 512 GB. A LocalMaxxing user reports about 33 tok/s for DeepSeek V4 Flash on a 512 GB M3 Ultra, and another reports 14.8 tok/s for Qwen3.8-Flash-Next.
 
 
Can you still buy a Mac Studio M4 Max or M3 Ultra?
 
Not new from Apple. On 3 November 2026 Apple’s US store sold only the M5 Max and M5 Ultra Mac Studio, and its refurbished store listed only M5 Max units; Apple had already dropped the 512 GB M3 Ultra in March 2026. An M4 Max or M3 Ultra now means a reseller or a used unit.
 
 
M5 Max or M5 Ultra for AI?
 
The Mac Studio M5 Ultra has about 1.2 TB/s of bandwidth per Apple, roughly twice the M5 Max, and goes to 512 GB, so it holds DeepSeek V4.1 Flash. For models up to 128 GB the M5 Max costs far less: $5,399 against $10,799 for a 256 GB Ultra.