No event, no keynote — Apple dropped both machines by press release this morning. Every figure below comes from Apple’s newsroom and spec pages, and every speed number is estimated from published memory bandwidth via the LLMCheck model — the machines ship 22 September and nobody has measured one yet, including us.
What actually shipped
| Machine | Chip | Bandwidth | Unified memory | From |
|---|---|---|---|---|
| Mac mini | M6 (2nm) | 153 GB/s (16GB bins) / 170 GB/s | 16 / 24 / 32 GB | $899 |
| Mac mini | M5 Pro | 307 GB/s | 24 / 48 / 64 GB | $1,699 |
| Mac Studio | M5 Max | 460 (32-core GPU) / 614 GB/s (40-core) | 36 / 48 / 64 / 128 GB | $2,499 |
| Mac Studio | M5 Ultra (quad-die) | 1.2 TB/s | 96 / 256 / 512 GB | $5,499 |
Source: Apple newsroom and tech-specs pages, 25 August 2026. Note the bins: the $899 mini’s 16GB configurations run 153 GB/s, and the base 32-core-GPU M5 Max Studio runs 460 — the headline bandwidth on both machines requires the upgraded configuration.
Two details matter more than the marketing. First, the mini’s M5 Pro runs 307 GB/s — 12% more than the 273 of the M4 Pro and of the M5 Pro in the MacBook Pro. Second, the M5 Ultra is Apple’s first quad-die chip, and its 1.2 TB/s is 50% above the M3 Ultra that until today held the 512 GB crown.
Why bandwidth is the whole story
Token generation on Apple Silicon is memory-bound: for every token, the runtime streams essentially all active model weights through the GPU. Bandwidth in, tokens out. That is why one number per chip predicts so much — and why the same model, quantised the same way, produces this ladder:
| Machine | Bandwidth | Qwen3.8-27B (4-bit) est. |
|---|---|---|
| Mac mini M6 (24GB+) | 170 GB/s | ~8 tok/s |
| Mac mini M5 Pro | 307 GB/s | ~15 tok/s |
| Mac Studio M5 Max (40-core) | 614 GB/s | ~30 tok/s |
| Mac Studio M4 Ultra | 1,092 GB/s | ~51 tok/s |
| Mac Studio M5 Ultra | 1,228 GB/s | ~57 tok/s |
All estimated from the published bandwidth via the LLMCheck two-term model; the 27B is the current index #1. No measured Apple Silicon figures exist for these machines yet.
The machine that changes the ceiling: M5 Ultra, 512 GB
A 512 GB Mac is not new — the M3 Ultra Studio got there at 819 GB/s. What is new is 512 GB at 1.2 TB/s. The models that only just fit a half-terabyte Mac were exactly the ones starved by bandwidth, and they gain the most:
| Model | Fits in | M5 Ultra est. |
|---|---|---|
| DeepSeek V4 Flash (284B-A13B) | 256 GB | ~80 tok/s |
| Mistral Small 4 (119B MoE) | 96 GB | ~86 tok/s |
| Qwen3-Coder-Next (80B-A3B) | 96 GB | ~72 tok/s |
| GLM-4.5-Air (106B MoE) | 96 GB | ~61 tok/s |
| Inkling-Small (276B MoE) | 256 GB | ~45 tok/s |
| Qwen3-235B-A22B | 256 GB | ~37 tok/s |
| Llama 3.3 70B (dense) | 96 GB | ~24 tok/s |
| Llama 3.1 405B (dense) | 512 GB | ~4 tok/s |
Estimated: dense models via the bandwidth formula, MoE scaled from each model’s existing reference figure by the bandwidth ratio. The 405B row is the honest one — it loads, and at ~4 tok/s it demonstrates why dense frontier models remain impractical even here.
What to buy at each budget
- $899–$1,399 — Mac mini M6. Skip the 16 GB bins (153 GB/s and no headroom); the 24 GB or 32 GB configs run every strong small model — and at 32 GB, even the index #1 Qwen3.8-27B fits at 4-bit, at ~8 tok/s. Patient-user territory, but it runs.
- $1,699–$2,299 — Mac mini M5 Pro. The new value pick. 307 GB/s and up to 64 GB: 27–35B models at usable speeds (~15 tok/s for the 27B dense, ~27 for 35B-A3B MoE), and 64 GB opens GLM-4.5-Air territory.
- $2,499–$3,999 — Mac Studio M5 Max. Get the 40-core GPU (614 GB/s) and 128 GB: 70B dense at ~28 tok/s and DeepSeek V4 Flash at 2-bit around ~40. The 24/7 workhorse.
- $5,499+ — Mac Studio M5 Ultra. The frontier machine. 96 GB covers everything through 120B-class MoE fast; 256 GB adds V4 Flash at 4-bit (~80 tok/s est.) and the 235B Qwen; 512 GB is for half-terabyte-class experiments.
Should you buy the outgoing generation instead? Probably, if price matters more than ceiling. M4-generation Studios will be discounted from today, an M4 Ultra at 1,092 GB/s gives up only ~11% to the new Ultra, and a used M3 Ultra 512 GB remains the cheapest path to half-terabyte models. The machines that got genuinely better this cycle are the minis: $899 now buys a real local-AI computer, and 307 GB/s in a mini did not exist at any price.
Every model, ranked for the new machines
The index now covers all four: Mac mini M6, Mac mini M5 Pro, Mac Studio M5 Max and Mac Studio M5 Ultra — every catalog model that fits each configuration, ranked on the same figures as the leaderboard. The Mac Advisor has the new configs too.