Mac Studio M5 Max 128GB vs M5 Ultra 96GB: Which for Local AI?
Posted Video released
Mac Studio M5 Max 128GB vs M5 Ultra 96GB for local AI: Ultra for speed at 27B and under, Max for 122B-class MoE. Measured and estimated numbers kept apart.
Which Mac Studio is better for local AI: the M5 Max 128GB or the M5 Ultra 96GB? The answer is "speed or capacity." For running dense models of 27B or smaller fast, choose the Ultra 96GB. For running a 122B-class MoE model, choose the Max 128GB. This article rewrites a YouTube video and separates every number into what we measured, what reviewers measured, official specs, and calculations or estimates. The video is narrated in Korean.
The conclusion in 3 lines
- To run dense models of 27B or smaller fast, get the Mac Studio M5 Ultra 96GB. The Ultra is about 55 tokens/s (reviewer-measured) against about 30 tokens/s for the Max (calculated estimate), roughly 2x.
- To run one 100B-class MoE model, get the M5 Max 128GB. Qwen3.5 122B needs about 74GB (estimate), and the Ultra 96GB has a GPU limit of about 72GB (estimate), so it does not fit.
- For models over 150GB or for simultaneous use, get the Ultra 256GB. For solo chat and summaries, a subscription is cheaper.
This article uses four kinds of numbers
The GitSeonjeap channel does not own a Mac Studio. Every Studio number is a reviewer measurement, an official spec, or a calculation.
| Label | Meaning |
|---|---|
| Measured by us | Values we measured ourselves on a Mac mini M6 32GB |
| Reviewer-measured | Values measured on real hardware by BGR, Tom's Hardware and others |
| Official | Public specs from Apple and NVIDIA |
| Calculated / estimate | Bandwidth ÷ model size, the 75% memory rule |
Prices come from two points in time, so we list them separately. On the video's reference date (2026-09-27), in Korea, the Max 128GB was about KRW 9,390,000 (1TB basis; the base price plus the option price added together), and the Ultra 96GB was KRW 9,490,000. The prices shown by Apple Korea on 2026-10-30 are KRW 8,710,000 for the Max 128GB (512GB storage by default) and KRW 9,490,000 for the Ultra 96GB (1TB storage by default). The default storage differs, so the two values cannot be compared directly. The real question is whether you want 32GB more RAM or twice the bandwidth.
Specs and price difference between the two configurations
The two models share the same exterior and ports. The difference is inside. The comparison below is as of the video's reference date (2026-09-27).
| Item | M5 Max 128GB | M5 Ultra 96GB |
|---|---|---|
| Price (Korea) | About KRW 9,390,000 (video reference date 9/27, 1TB basis) · KRW 8,710,000 (official price shown on 10/30, 512GB storage by default) | KRW 9,490,000 |
| CPU / GPU | 18 cores / 40 cores | 30 cores / 64 cores |
| Memory | 128GB | 96GB |
| Memory bandwidth | 614GB/s | 1,200GB/s |
| Maximum memory | 128GB | 512GB (expected late October as of the video) |
Speed comes from bandwidth, model size comes from memory
The speed at which text is generated is roughly proportional to memory bandwidth. In the video's analogy, capacity is the size of the warehouse and bandwidth is the width of its door.
Our reference point is what we measured directly on a Mac mini M6 32GB (MLX 4-bit, 2026-09-26).
Qwen3.8 27B 4-bit reached 9.6 tokens/s, Qwen3.6 35B-A3B reached 67.4 tokens/s, and gpt-oss 20B reached 46.0 tokens/s (all measured by us).
Here is the same Qwen3.8 27B 4-bit lined up across machines.
The Mac mini M6 is 9.6 (measured by us). The M5 Max is about 30 (estimate calculated from bandwidth; no measurement. The formula is not recorded in the video's materials, so treat it as a rounded value. For reference, scaling the Ultra's 55 down by the bandwidth ratio gives 55×614÷1,200≈28, and scaling the M6's 9.6 up gives 9.6×614÷170≈35; calculated). The M3 Ultra is about 40 and the M5 Ultra about 55 (both reviewer-measured by BGR; the Ultra was the 256GB configuration).
Bandwidth differs by 7.1x (170 vs 1,200), but speed differs by only about 5.7x (9.6 vs 55). The video explains that a larger chip cannot use its full bandwidth, that prompt reading depends on GPU core count, and that MLX is less optimized for the new chips. The machines differ, so the 5.7x figure is a reference calculation only.
How large a model each memory size can run
A Mac does not give all of its memory to the GPU. On our M6, the limit was 26.8GB (24.96GiB) out of 32GB (measured by us). Counting 32GB as 32GiB (about 34.4GB), that is 78%. Dividing by 32 in decimal GB (10^9 bytes) gives about 84% (26.8÷32, calculated). We could not measure the Studio the same way, so we used an estimate with a conservative 75% applied. That 75% is multiplied directly by the GB number in each configuration's name, so we did not separate the units.
At these limits, Qwen3.5 122B-A10B 4-bit decides the matter. It needs about 74GB (estimate): about 68GB for the model plus about 6GB for a 32k-token context cache. It fits in the Max 128GB and does not fit in the Ultra 96GB.
By contrast, a 70B dense 4-bit model (about 40GB) fits in both, and bandwidth sets the speed. Calculated from bandwidth, the Max reaches about 12 tokens/s and the Ultra about 22 tokens/s (estimate, not measured). The video's materials do not record the efficiency factor that was applied. Working backward, the values are close to bandwidth ÷ 40GB multiplied by 75% (614÷40×0.75=11.5, 1,200÷40×0.75=22.5; calculated). Multiplying by 80% gives 12.3 and 24 (calculated).
| Configuration | Price (Korea) | Bandwidth (official) | GPU limit (estimate) | Model sizes it can run |
|---|---|---|---|---|
| M5 Ultra 96GB | KRW 9,490,000 | 1,200GB/s | About 72GB | 27B and 31B dense models, 35B-A3B-class MoE |
| M5 Max 128GB | KRW 8,710,000 (10/30, 512GB storage by default) | 614GB/s | About 96GB | Everything above, plus 122B-class MoE |
| M5 Ultra 256GB | KRW 16,290,000 (the 9/27 video value; the same in the Apple Korea configurator on 10/30) | 1,200GB/s | About 192GB | DeepSeek V4 Flash 156GB (reviewer-measured by ZONEofTECH) |
Qwen3.5 122B on the Ultra 256GB runs at about 80 tokens/s, reviewer-measured by BGR.
How many months of subscription is that?
We converted the prices into a subscription such as ChatGPT Pro or Claude Max at about KRW 290,000 per month (based on prices on the video's reference date, excluding electricity and depreciation; calculated). The Ultra 96GB equals 33 months, the Max 128GB (based on about KRW 9,390,000) equals 32 months, and the Ultra 256GB equals 56 months. In years, that is about 2.7 to 4.7 (32÷12 and 56÷12; calculated), which the video rounded to "3 to 5 years."
Here is how it compared with NVIDIA. In Tom's Hardware's measurements, the M5 Ultra generated tokens about 4x faster than the DGX Spark (128GB, bandwidth 273). The RTX 5090 (32GB, bandwidth 1,792) is faster for models that fit, but it is limited to 32GB.
Order of choices before you buy
These are the video's verdicts, in order.
- Decide the model size first. Dense models of 27B or smaller and coding assistance fit the M5 Ultra 96GB.
- One 100B-class MoE model and long context fit the M5 Max 128GB.
- Frontier-class MoE models over 150GB, or simultaneous use, fit the Ultra 256GB.
- If solo chat and summaries are all you do, a subscription (about KRW 290,000 per month) is cheaper.
The video named three cases where buying makes sense: when documents must not leave your machine, when you run an agent 24 hours a day, and when simultaneous use makes token fees a burden. As of the video, delivery in Korea took 10 to 12 weeks (cited from a Clien post, 2026-09-14). The 3x figure for a 4-unit cluster is Apple's claim.
Summary
- Speed goes to the Ultra 96GB (KRW 9,490,000), and capacity goes to the Max 128GB. The Max price differs by date: about KRW 9,390,000 on the video's reference date, and KRW 8,710,000 shown on 2026-10-30 (512GB storage by default).
- The Studio numbers are reviewer measurements and calculated estimates. The only numbers we measured ourselves are for the Mac mini M6.
Sources, reference date and limits
- Reference date: the video was published 2026-09-27. We updated prices to the official prices shown by Apple Korea on 2026-10-30 (before discounts). Stock, delivery and schedules remain as of the video and may have changed. The 512GB configuration was expected in late October as of the video's reference date. In the Apple Korea configurator on 2026-10-30, it appeared as an option on the 36-core CPU / 80-core GPU Ultra, priced at KRW 38,050,000.
- Sources: Apple Newsroom (2026-08-25, specs), BGR review (2026-09-21), Tom's Hardware review (2026-09-21), Context Studios, ZONEofTECH review. The Max price on the video's reference date of September 27 (about KRW 9,390,000) is the base price plus option prices added together. The values from October 30 have different default storage (Max 512GB, Ultra 1TB), so they are not a comparison at matched storage.
- Limits: we could not measure a real Mac Studio. The Max's roughly 30 tokens/s for 27B, the 70B speeds and the GPU limits are calculated estimates. The BGR numbers came from a 256GB configuration (described in the video as having an 80-core GPU), so they may differ from the base 96GB model with a 64-core GPU.
This article is based on a video narrated with an AI voice, and AI assisted the writing. This is not a purchase recommendation.