The verdict is the Mac mini M5 Pro (48 GB), if the $1,000 premium fits: the LLMCheck index’s figures put it a median 1.7× faster across the 61 shared models with a figure on both; on Qwen 3.6-35B-A3B, 79 against 59 tok/s (both community runs). It also fits 8 more of the index’s models.
| M6 (32 GB) | M5 Pro (48 GB) | |
|---|---|---|
| Chip | M6 | M5 Pro |
| GPU cores | 12-core | 16- or 20-core |
| Memory bandwidth | 170 GB/s | 307 GB/s |
| Unified memory | 32 GB | 48 GB |
| Price | $1,299 | $2,299 |
| Price per GB | $41 | $48 |
| Models that fit (of 130) | 63 | 71 |
| Best local LLM here | Qwen 3.6-35B-A3B community59 tok/s | LFM2.5-24B-A2B community114 tok/s |
| Qwen 3.6-35B-A3B | community59 tok/s | community79 tok/s |
| Price per tok/s on Qwen 3.6-35B-A3B | $22 | $29 |
| Where to buy | New · build-to-order from Apple | New · build-to-order from Apple |
Every speed below is the LLMCheck index’s figure for that exact configuration — the same number its Mac mini M6 (32 GB) and Mac mini M5 Pro (48 GB) pages print, with its provenance mark. Across the 61 shared models with a figure on both, the M5 Pro column is a median ×1.71 of the M6 column.
| Model | Size | M6 tok/s | M5 Pro tok/s | Ratio |
|---|---|---|---|---|
| Qwen 3.6-35B-A3B | 35B MoE | community59 | community79 | ×1.34 |
| Qwen3.8-27B | 27.8B | community10 slow | sourced17 | ×1.70 |
| Nanbeige4.3-33B | 33.1B | estimated7 slow | estimated12 slow | ×1.71 |
| Qwen 3.6-27B | 27B | estimated9 slow | estimated15 | ×1.67 |
| KAT-Coder-V2.5 | 35B MoE | estimated47 | estimated72 | ×1.53 |
| LFM2.5-24B-A2B | 24B MoE | estimated79 | community114 | ×1.44 |
| Recursive Flint | 39.8B MoE | estimated61 | community92 | ×1.51 |
| Muse Glimmer 30B | 30B | community9 slow | estimated11 slow | ×1.22 |
| MiniCPM-V 4.7 35B-A3B | 35.2B MoE | estimated54 | community57 | ×1.06 |
| Nemotron 3.5 Lightning | 30B MoE | community45 | community72 | ×1.60 |
| Gemma 4 31B | 31B | community8 slow | estimated13 slow | ×1.62 |
| Raon-1 mini preview | 28.8B MoE | estimated58 | community86 | ×1.48 |
| Gemma 4 26B-A4B | 26B MoE | community49 | estimated78 | ×1.59 |
| GLM-4.7-Flash | 31B MoE | community47 | estimated73 | ×1.55 |
| North Mini Code 1.0 | 30.5B MoE | estimated37 | community70 | ×1.89 |
A model fits when its minimum RAM (its smallest build of at least 3 bits with a working Mac runtime, plus about 2 GB for the runtime and a short context; 2-bit only where a published Mac run used it) is at most the Mac’s unified memory. The extra memory unlocks these models from the LLMCheck index:
| Model | Size | Min RAM | M5 Pro tok/s |
|---|---|---|---|
| Nemotron 3.5 Bolt 74B-A9B caveat | 74B MoE | 45 GB | estimated31 |
| Ripple-65B Preview | 64.8B | 33 GB | estimated7 |
| Qwen3-Coder-Next | 80B MoE | 48 GB | estimated49 |
| DeepSeek R1 70B | 70B | 44 GB | estimated6 |
| Apertus 1.5 70B | 70B | 44 GB | estimated6 |
| Hermes 4 70B | 70B | 38 GB | estimated6 |
| Qwen 2.5 72B | 72B | 46 GB | estimated6 |
| Llama 3.3 70B | 70B | 44 GB | estimated6 |
At Apple’s list prices the M6 (32 GB) costs $41 per GB of unified memory and the M5 Pro (48 GB) $48. On Qwen 3.6-35B-A3B — the reference model here, which both run — that is $22 and $29 per tok/s of decode speed: the M6 (32 GB) buys more speed per dollar.
The $1,000 premium (77% more) buys +34% decode speed on Qwen 3.6-35B-A3B and 8 more models that fit.
Bottom line: The M5 Pro (48 GB): a median 1.7× the speed for $1,000 more.
Amazon does not stock either configuration: Apple builds both to order.
Each speed is decode (generation) tok/s from the LLMCheck index for that exact configuration: the lower-median measured run on that chip where one exists (a filled square is a published figure, a hollow circle a community run), otherwise an estimate from the memory-bandwidth model (hollow diamond, printed with a tilde). LLMCheck runs no tests of its own. Prompt processing (prefill) depends on GPU compute rather than bandwidth, and the index does not estimate it.
The verdict is the Mac mini M5 Pro (48 GB), if the $1,000 premium fits: the LLMCheck index’s figures put it a median 1.7× faster across the 61 shared models with a figure on both; on Qwen 3.6-35B-A3B, 79 against 59 tok/s (both community runs). It also fits 8 more of the index’s models.
Decode speed on Apple Silicon follows memory bandwidth: the Mac mini M5 Pro (48 GB) reads memory at 307 GB/s, the Mac mini M6 (32 GB) at 170 GB/s. The LLMCheck index puts Qwen 3.6-35B-A3B at 79 tok/s (community run) against 59 tok/s (community run). Across the 61 shared models with a figure on both, the median gain is 1.71x.
8 models in the LLMCheck index need more than 32 GB but fit in 48 GB, led by Nemotron 3.5 Bolt 74B-A9B, Ripple-65B Preview, Qwen3-Coder-Next. Both run the other 63.
It costs $2,299 against $1,299 (Apple list prices, checked 14 Nov 2026; the M5 Pro (48 GB) price includes 512GB of storage, the M6 (32 GB) price 256GB). For that it gives 79 tok/s (community run) against 59 tok/s (community run) on Qwen 3.6-35B-A3B and fits 8 more models. Worth it if you use the extra speed or models daily; the cheaper Mac still runs all but 8 of the same models.
On the Mac mini M6 (32 GB): Qwen 3.6-35B-A3B at 59 tok/s (community run). On the Mac mini M5 Pro (48 GB): LFM2.5-24B-A2B at 114 tok/s (community run). These are the headline picks of the LLMCheck Best by Mac pages: the best-ranked model that reaches 15 tok/s on that Mac and has a standard Mac runtime.