The shortlist: used Macs for local LLMs
Each row is one configuration worth seeking out, the largest kind of model it holds, that model’s decode speed on it, and Qwen 3.6-35B-A3B, one of the index’s two top-ranked models, for comparison. Every speed is filled from the LLMCheck index at build time and carries its provenance; on these older chips most are estimates from the published bandwidth model. Each Mac links to its Best-by-Mac page, which ranks everything that fits.
| Used Mac | Bandwidth | Largest class it holds | Its speed | Qwen 3.6-35B-A3B | Buy renewed |
|---|---|---|---|---|---|
| Mac Studio M1 Ultra, 64 GB | 800 GB/s | Llama 3.3 70B | estimated~15 tok/s | estimated~105 tok/s | M1 Ultra Renewed on Amazon |
| Mac Studio M1 Ultra, 128 GB | 800 GB/s | Qwen3.8-Flash-Next | estimated~57 tok/s | estimated~105 tok/s | M1 Ultra 128GB Renewed on Amazon |
| Mac Studio M2 Ultra, 192 GB | 800 GB/s | MiMo-V2.6-Flash | estimated~54 tok/s | estimated~105 tok/s | M2 Ultra Renewed on Amazon |
| Mac Studio M2 Max, 96 GB | 400 GB/s | GPT-oss 120B | estimated~66 tok/s | estimated~92 tok/s | M2 Max Renewed on Amazon |
| MacBook Pro M3 Max, 128 GB | 400 GB/s | Qwen3.8-Flash-Next | estimated~35 tok/s | estimated~92 tok/s | M3 Max Renewed on Amazon |
| Mac Studio M4 Max, 128 GB | 546 GB/s | Qwen3.8-Flash-Next | estimated~45 tok/s | estimated~105 tok/s | Mac Studio M4 Max Renewed on Amazon |
| Mac Studio M3 Ultra, 256–512 GB | 819 GB/s | MiMo-V2.6-Flash | community59 tok/s | community106 tok/s | M3 Ultra Renewed on Amazon |
| Mac mini M4 Pro, 48–64 GB | 273 GB/s | Qwen3-Coder-Next | estimated~44 tok/s | estimated~74 tok/s | Mac mini M4 Pro Renewed on Amazon |
Prices move weekly and depend on condition, storage and seller, so this guide quotes none it cannot date: check the current price on the listing. For scale, Apple’s new equivalents are a Mac Studio M5 Max with 128 GB at $5,099 and an M5 Ultra with 256 GB at $9,499 (Apple list prices, checked 14 November 2026).
M1 Ultra and M2 Ultra: 800 GB/s for the price of a new mid-range Mac
The Ultra Mac Studios of 2022 and 2023 have 800 GB/s of memory bandwidth, more than any new Mac except the M5 Ultra (1,228 GB/s); the M5 Max has 614. On paper that out-reads a new M5 Max on large dense models, where decode follows bandwidth most closely: the index estimates Llama 3.3 70B at estimated~15 tok/s on an M1 Ultra against sourced13 tok/s on an M5 Max. But no M1 or M2 Ultra run of a large model is in the index yet, and in llama.cpp’s own Apple Silicon table (Llama 2 7B) the 64-core M1 Ultra only matches an M5 Max at 16-bit and trails it at 4-bit, while the 76-core M2 Ultra leads only at 16-bit: expect parity on large dense models, not a lead. On small mixture-of-experts models the newer chip closes or reverses the gap (Qwen 3.6-35B-A3B: estimated~105 tok/s against community130 tok/s), and either is far faster than a new Mac mini.
- M1 Ultra (Mac Studio, 2022): 64 or 128 GB, 48- or 64-core GPU, both at 800 GB/s. At 64 GB it holds 70B-class dense models at 4-bit; at 128 GB, 120B-class mixture-of-experts models such as Qwen3.8-Flash-Next and DeepSeek V4 Flash at 2-bit.
- M2 Ultra (Mac Studio and Mac Pro, 2023): 64, 128 or 192 GB, 60- or 76-core GPU, 800 GB/s. The 192 GB model is the cheapest way to hold MiMo-V2.6-Flash, the index’s pick from 192 GB up, which needs about 170 GB: estimated~54 tok/s on it.
Two limits. Neither generation has the Neural Accelerators that speed up prompt processing on M5 chips, so long prompts (coding agents, RAG over many documents) wait longer before the first token than on a new Mac with the same bandwidth. And both have Thunderbolt 4, not Thunderbolt 5, so they cannot join the RDMA clusters covered in the multi-Mac guide.
M2 Max 96 GB, M3 Max 128 GB, M4 Max 128 GB
The Max chips are where used laptops come in, and the Mac Studio versions often sell for less than the MacBook Pros.
- M2 Max with 96 GB (MacBook Pro and Mac Studio, 2023; the 38-core GPU only): 400 GB/s. It holds GPT-oss 120B, which needs about 62 GB, at estimated~66 tok/s, and Qwen3-Coder-Next (48 GB) at estimated~61 tok/s.
- M3 Max with 128 GB (MacBook Pro, late 2023; 40-core GPU): 400 GB/s. It holds the 120B-class models: Qwen3.8-Flash-Next at estimated~35 tok/s, DeepSeek V4 Flash at 2-bit at community27 tok/s.
- M4 Max with 128 GB (MacBook Pro, late 2024; Mac Studio, 2025; 40-core GPU): 546 GB/s, the fastest of the three and the only one with Thunderbolt 5. Qwen3.8-Flash-Next runs at estimated~45 tok/s, Qwen 3.6-35B-A3B at estimated~105 tok/s.
Below these, a used Mac mini M4 Pro with 48 or 64 GB (273 GB/s, Thunderbolt 5) is the small-desk option: it holds the same 30B-class mixture-of-experts models as a new Mac mini M5 Pro with 11% less bandwidth, and Qwen3-Coder-Next at 64 GB. A used M3 Ultra Mac Studio (819 GB/s) is the high end: Apple stopped selling the 512 GB option around 5 March 2026 and the model on 25 August, and at 512 GB it holds DeepSeek V4.1 Flash at community20 tok/s.
The variants that cost you bandwidth
Listings name the chip, and the chip name hides the variant. These are the cases where the same name means less bandwidth:
| Listing says | Variant to avoid | Bandwidth | The one to buy |
|---|---|---|---|
| M3 Max | 14-core CPU, 30-core GPU (sold with 36 or 96 GB) | 300 GB/s | 16-core CPU, 40-core GPU: 400 GB/s, 48–128 GB |
| M4 Max | 14-core CPU, 32-core GPU (36 GB only) | 410 GB/s | 16-core CPU, 40-core GPU: 546 GB/s, 48–128 GB |
| M2 Max 96 GB | none: 96 GB came only with the 38-core GPU | 400 GB/s | any 96 GB unit |
| M1 / M2 Ultra | none: the smaller GPU has the same 800 GB/s | 800 GB/s | the most memory you can afford |
The 96 GB M3 Max is the trap that matters: it has more memory than a 64 GB 40-core M3 Max but a quarter less bandwidth, so everything that fits both runs slower on it. Check the GPU core count in the listing, or in About This Mac → More Info → System Report → Graphics/Displays before you pay.
Where to buy: Amazon Renewed and Apple Certified Refurbished
Apple Certified Refurbished is the safest source: Apple says each unit gets “full functional testing”, genuine replacement parts where needed and its standard one-year limited warranty, and AppleCare can be added (Apple’s refurbished page). Stock is thin for high-memory Macs: on 18 November 2026 Apple’s refurbished Mac store listed no Mac Studio at all (its Mac Studio page now redirects there). LLMCheck earns nothing from Apple; Apple has no affiliate program for Macs.
Amazon Renewed listings come from third-party refurbishers, and condition, battery health and warranty vary by seller, so read the listing’s exact configuration and the seller’s terms. The searches below open Amazon’s Renewed results for each machine:
Renewed high-memory Macs on Amazon
The configurations to look for are in the shortlist above. Check the memory, the GPU core count and the seller’s return window on every listing.
Mac Studio M2 Ultra Renewed on Amazon → Mac Studio M1 Ultra Renewed on Amazon → Mac Studio M4 Max Renewed on Amazon → MacBook Pro M3 Max Renewed on Amazon →Disclosure: As an Amazon Associate, LLMCheck earns from qualifying purchases. This doesn’t affect the index’s figures or rankings.
Private sales and marketplaces are cheaper still and carry no warranty; the check below matters most there.
Before you pay: a seven-point check
- Memory and GPU cores match the listing (System Report → Hardware and Graphics/Displays). Unified memory cannot be upgraded later.
- Activation Lock is off: the seller has signed out of their Apple Account and erased the Mac.
- Battery cycle count and condition on a laptop (System Report → Power). Long local-model sessions are hard on a battery that is already worn.
- The fans spin up under load. Run a model for a few minutes; a long generation is a fair stress test.
- macOS 26 installs. Every Apple-silicon Mac runs it, and current Ollama loads its newest MLX build only on macOS 26 (Ollama MLX guide).
- Storage is big enough. A 120B-class model is 60–120 GB on disk per quant; 1 TB fills fast.
- The return window is long enough to run your own model on it.
Then load the model you bought it for and compare its decode speed with the figure on that Mac’s Best-by-Mac page. If you can, submit the run: community runs are how the estimates on these older chips get replaced.
Related guides
- Best Mac for local LLMs: every new Mac compared, with prices.
- Mac Advisor: set a budget, allow refurbished, and get one configuration.
- Apple Silicon chips compared: bandwidth and speed for every chip from M1 to M6.
- Best LLM by RAM tier: what each memory size runs.