For models above 128 GB the verdict is a used Mac Studio M3 Ultra (256 GB): it fits 15 more of the index's models, led by MiMo-V2.6-Flash at 59 tok/s (community run). For everything that fits in 128 GB the new M5 Max (128 GB) is a little slower and has a warranty.
| M3 Ultra (256 GB) | M5 Max (128 GB) | |
|---|---|---|
| Chip | M3 Ultra | M5 Max |
| GPU cores | 60- or 80-core | 40-core |
| Memory bandwidth | 819 GB/s | 614 GB/s |
| Unified memory | 256 GB | 128 GB |
| Price | — | $5,099 |
| Price per GB | — | $40 |
| Models that fit (of 130) | 108 | 93 |
| Best local LLM here | MiMo-V2.6-Flash community59 tok/s | Qwen 3.6-35B-A3B community130 tok/s |
| Qwen3.8-Flash-Next | estimated57 tok/s | community47 tok/s |
| Price per tok/s on Qwen3.8-Flash-Next | — | $108 |
| Where to buy | Used / refurbished (Apple no longer sells it new) | 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 M3 Ultra (256 GB) and Mac Studio M5 Max (128 GB) pages print, with its provenance mark. Across the 89 shared models with a figure on both, the M5 Max column is a median ×0.80 of the M3 Ultra column.
| Model | Size | M3 Ultra tok/s | M5 Max tok/s | Ratio |
|---|---|---|---|---|
| Qwen 3.6-35B-A3B | 35B MoE | community106 | community130 | ×1.23 |
| Qwen3.8-Flash-Next | 125B MoE | estimated57 | community47 | ×0.82 |
| Recursive Flint | 39.8B MoE | community91 | community123 | ×1.35 |
| DeepSeek V4 Flash | 284B MoE | estimated32 | community39 | ×1.22 |
| DeepSeek V4 Flash Vision-Exp | 305B MoE | estimated32 | estimated39 | ×1.22 |
| GLM-5.3-Flash | 320B MoE | estimated24 | estimated27 | ×1.12 |
| Inkling-Small caveat | 276B MoE | estimated35 | estimated42 | ×1.20 |
| North Mini Code 1.0 | 30.5B MoE | estimated133 | community120 | ×0.90 |
| Nemotron 3.5 Lightning | 30B MoE | estimated151 | community117 | ×0.77 |
| LFM2.5-24B-A2B | 24B MoE | community116 | sourced184 | ×1.59 |
| Qwen3.8-27B | 27.8B | community38 | community31 | ×0.82 |
| KAT-Coder-V2.5 | 35B MoE | estimated100 | community89 | ×0.89 |
| Qwen3-235B-A22B | 235B MoE | estimated32 | estimated26 | ×0.81 |
| Nanbeige4.3-33B | 33.1B | estimated35 | community25 | ×0.71 |
| Raon-1 mini preview | 28.8B MoE | estimated136 | estimated123 | ×0.90 |
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 | M3 Ultra tok/s |
|---|---|---|---|
| MiMo-V2.6-Flash | 309B MoE | 170 GB | community59 |
| Inkling 2 Small caveat | 594B MoE | 228 GB | estimated36 |
| GLM 5.2 | 753B MoE | 256 GB | estimated15 |
| MiniMax-M3 | 428B MoE | 162 GB | estimated29 |
| GLM-5.3 | 753B MoE | 256 GB | estimated11 |
| MiniMax-M3.1-Flash | 428B MoE | 189 GB | community26 |
| Beam 1.0 caveat | 501.3B MoE | 199 GB | estimated19 |
| MiniMax M2.5 | 230B MoE | 130 GB | estimated41 |
| Hunyuan Hy3 | 295B MoE | 160 GB | estimated21 |
| Nemotron 3 Ultra 550B-A55B | 550B MoE | 226 GB | estimated12 |
The new M5 Max (128 GB) costs $5,099 (Apple list, 14 Nov 2026): $40 per GB and $108 per tok/s on Qwen3.8-Flash-Next. A used M3 Ultra (256 GB) matches that $108 per tok/s at about $6,200; it decodes faster, so any used price below that buys more speed per dollar than the new Mac. Its real case is the 15 models that need more than 128 GB, which the new Mac cannot load at any price. The index records no used prices: check current listings, and weigh the warranty and the remaining years of macOS support.
Bottom line: A used M3 Ultra (256 GB) for the 15 models above 128 GB; the new M5 Max (128 GB) for the rest.
Renewed stock of the 256 GB Mac Studio M3 Ultra varies: check the memory in the listing before you buy. Amazon does not stock the Mac Studio M5 Max (128 GB): Apple builds it to order.
As an Amazon Associate, LLMCheck earns from qualifying purchases. Affiliate links cost you nothing extra and never influence the verdicts.
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; Apple’s M5-generation GPUs add per-core Neural Accelerators aimed at that work.
For models above 128 GB the verdict is a used Mac Studio M3 Ultra (256 GB): it fits 15 more of the index's models, led by MiMo-V2.6-Flash at 59 tok/s (community run). For everything that fits in 128 GB the new M5 Max (128 GB) is a little slower and has a warranty.
Decode speed on Apple Silicon follows memory bandwidth: the Mac Studio M3 Ultra (256 GB) reads memory at 819 GB/s, the Mac Studio M5 Max (128 GB) at 614 GB/s. The LLMCheck index puts Qwen3.8-Flash-Next at ~57 tok/s (estimated) against 47 tok/s (community run). Across the 89 shared models with a figure on both, the median gain is 1.25x.
15 models in the LLMCheck index need more than 128 GB but fit in 256 GB, led by MiMo-V2.6-Flash, Inkling 2 Small, GLM 5.2. Both run the other 93.
About $6,200 or less. At that price it matches the new Mac's $108 per tok/s on Qwen3.8-Flash-Next ($5,099 at Apple list, checked 14 Nov 2026). The LLMCheck index records no used prices, so check current listings, and weigh warranty and macOS support.
On the Mac Studio M3 Ultra (256 GB): MiMo-V2.6-Flash at 59 tok/s (community run). On the Mac Studio M5 Max (128 GB): Qwen 3.6-35B-A3B at 130 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.