For most local-LLM work the verdict is the Mac Studio M5 Ultra (96 GB): a median 1.7× faster than the M5 Max (128 GB) across the 82 shared models with a figure on both; on Qwen 3.6-35B-A3B, 153 against 130 tok/s (both community runs). The M5 Max (128 GB) fits 7 more models.
| M5 Max (128 GB) | M5 Ultra (96 GB) | |
|---|---|---|
| Chip | M5 Max | M5 Ultra |
| GPU cores | 40-core | 64- or 80-core |
| Memory bandwidth | 614 GB/s | 1,228 GB/s |
| Unified memory | 128 GB | 96 GB |
| Price | $5,099 | $5,499 |
| Price per GB | $40 | $57 |
| Models that fit (of 130) | 93 | 86 |
| Best local LLM here | Qwen 3.6-35B-A3B community130 tok/s | KAT-Coder-V2.5 estimated120 tok/s |
| Qwen 3.6-35B-A3B | community130 tok/s | community153 tok/s |
| Price per tok/s on Qwen 3.6-35B-A3B | $39 | $36 |
| 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 Studio M5 Max (128 GB) and Mac Studio M5 Ultra (96 GB) pages print, with its provenance mark. Across the 82 shared models with a figure on both, the M5 Ultra column is a median ×1.68 of the M5 Max column.
| Model | Size | M5 Max tok/s | M5 Ultra tok/s | Ratio |
|---|---|---|---|---|
| KAT-Coder-V2.5 | 35B MoE | community89 | estimated120 | ×1.35 |
| Qwen 3.6-35B-A3B | 35B MoE | community130 | community153 | ×1.18 |
| Ling-3.0-flash-VL | 124B MoE | estimated53 | estimated89 | ×1.68 |
| Nanbeige4.3-33B | 33.1B | community25 | estimated52 | ×2.08 |
| Recursive Flint | 39.8B MoE | community123 | estimated131 | ×1.07 |
| Inkling-Small caveat | 276B MoE | estimated42 | estimated49 | ×1.17 |
| Qwen3.8-27B | 27.8B | community31 | sourced48 | ×1.55 |
| Ling-3.0 Flash | 124B MoE | community68 | estimated114 | ×1.68 |
| Laguna S 2.1 | 118B MoE | estimated38 | estimated68 | ×1.79 |
| Nemotron 3.5 Lightning | 30B MoE | community117 | estimated165 | ×1.41 |
| North Mini Code 1.0 | 30.5B MoE | community120 | estimated150 | ×1.25 |
| Nemotron 3.5 Super VL 120B-A12B | 120B MoE | community36 | community74 | ×2.06 |
| Qwen 3.6-27B | 27B | estimated30 | estimated57 | ×1.90 |
| Kolibri-1 | 78.1B MoE | estimated75 | estimated120 | ×1.60 |
| GLM-4.7-Flash | 31B MoE | estimated107 | estimated140 | ×1.31 |
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 Max tok/s |
|---|---|---|---|
| Qwen3.8-Flash-Next | 125B MoE | 120 GB | community47 |
| DeepSeek V4 Flash | 284B MoE | 120 GB | community39 |
| DeepSeek V4 Flash Vision-Exp | 305B MoE | 120 GB | estimated39 |
| GLM-5.3-Flash | 320B MoE | 128 GB | estimated27 |
| Qwen3-235B-A22B | 235B MoE | 128 GB | estimated26 |
| dots3-note preview | 280B MoE | 120 GB | estimated28 |
| Step-3.5-Flash | 196B MoE | 120 GB | estimated29 |
At Apple’s list prices the M5 Max (128 GB) costs $40 per GB of unified memory and the M5 Ultra (96 GB) $57. On Qwen 3.6-35B-A3B — the reference model here, which both run — that is $39 and $36 per tok/s of decode speed: the M5 Ultra (96 GB) buys more speed per dollar.
The $400 premium (8% more) buys +18% decode speed on Qwen 3.6-35B-A3B, but 7 fewer.
Bottom line: The M5 Ultra (96 GB) for speed (a median 1.7×); the M5 Max (128 GB) for 7 more models.
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.
For most local-LLM work the verdict is the Mac Studio M5 Ultra (96 GB): a median 1.7× faster than the M5 Max (128 GB) across the 82 shared models with a figure on both; on Qwen 3.6-35B-A3B, 153 against 130 tok/s (both community runs). The M5 Max (128 GB) fits 7 more models.
Decode speed on Apple Silicon follows memory bandwidth: the Mac Studio M5 Ultra (96 GB) reads memory at 1,228 GB/s, the Mac Studio M5 Max (128 GB) at 614 GB/s. The LLMCheck index puts Qwen 3.6-35B-A3B at 153 tok/s (community run) against 130 tok/s (community run). Across the 82 shared models with a figure on both, the median gain is 1.68x.
7 models in the LLMCheck index need more than 96 GB but fit in 128 GB, led by Qwen3.8-Flash-Next, DeepSeek V4 Flash, DeepSeek V4 Flash Vision-Exp. Both run the other 86.
It costs $5,499 against $5,099 (Apple list prices, checked 14 Nov 2026; the M5 Ultra (96 GB) price includes 1TB of storage, the M5 Max (128 GB) price 512GB). For that it gives 153 tok/s (community run) against 130 tok/s (community run) on Qwen 3.6-35B-A3B, while fitting 7 fewer. Worth it if you use the extra speed or models daily; the cheaper Mac still runs every model the dearer one does.
On the Mac Studio M5 Max (128 GB): Qwen 3.6-35B-A3B at 130 tok/s (community run). On the Mac Studio M5 Ultra (96 GB): KAT-Coder-V2.5 at ~120 tok/s (estimated). 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.