Compare two machines for local AI
Pick any two Macs, iPhones or GPUs. See the best model each one runs, how fast, and how much context fits in memory, from the same engine as the homepage.
12-core CPU, 12-core GPU · 170 GB/s
18-core CPU, 32-core GPU · 460 GB/s (the only M5 Max sold with 36GB)
Verdict at 32k context
The Mac mini M6 32GB tops out at Qwen3.8 27B (27B, ~8 tok/s est.); the Mac Studio M5 Max 36GB steps up to Qwen3.5 35B-A3B Instruct (35B MoE, 3B active, ~59 tok/s est.). On the same Qwen3.8 27B, the Mac Studio M5 Max 36GB is 2.7x faster.
Mac mini M6 32GB
BEST PICK AT 32K CONTEXT
$ ollama run qwen3.8:27b
Mac Studio M5 Max 36GB
BEST PICK AT 32K CONTEXT
$ ollama run qwen3.5:35b-a3b
Same model, both machines
Each machine's best pick, run on the other one too, at 32k context.
| Model | Mac mini M6 32GB | Mac Studio M5 Max 36GB | Difference |
|---|---|---|---|
| Qwen3.8 27B best pick of Mac mini M6 32GB | Fits ~8 tok/s est. | Fits ~22 tok/s est. | 2.7x faster on Mac Studio M5 Max 36GB |
| Qwen3.5 35B-A3B Instruct best pick of Mac Studio M5 Max 36GB | Fits ~22 tok/s est. | Fits ~59 tok/s est. | 2.7x faster on Mac Studio M5 Max 36GB |
Head to head
The top local models for either machine, biggest first. "Fits" means weights plus a 32k KV cache stay inside the memory budget.
| Model | Weights | Mac mini M6 32GB | Mac Studio M5 Max 36GB |
|---|---|---|---|
| Qwen3.6 35B-A3B 35B MoE, 3B active | 22 GB | Tight fit ~20 tok/s est. | Fits ~59 tok/s est. |
| Laguna XS 2.1 33B MoE, 3B active | 20.3 GB | Tight fit ~22 tok/s est. | Tight fit ~61 tok/s est. |
| Qwen3.5 35B-A3B Instruct 35B MoE, 3B active | 20 GB | Fits ~22 tok/s est. | Fits ~59 tok/s est. |
| GLM-4.7-Flash 30B-A3B 30B MoE, 3B active | 19 GB | Tight fit ~24 tok/s est. | Tight fit ~64 tok/s est. |
| Qwen3.6 27B 27B | 18 GB | Fits ~8 tok/s est. | Fits ~22 tok/s est. |
| Qwen3.8 27B 27B | 16.5 GB | Fits ~8 tok/s est. | Fits ~22 tok/s est. |
| Gemma 4 26B-A4B 26B MoE, 4B active | 16 GB | Tight fit ~22 tok/s est. | Fits ~59 tok/s est. |
| Qwen3.5 27B Instruct 27B | 16 GB | Fits ~8 tok/s est. | Fits ~22 tok/s est. |
| LFM2 24B-A2B Instruct 24B MoE, 2B active | 14 GB | Fits ~32 tok/s est. | Fits ~87 tok/s est. |
| GPT-OSS 20B 21B MoE, 3.6B active | 13.8 GB | Fits ~27 tok/s est. | Fits ~73 tok/s est. |
The hardware behind the numbers
| Spec | Mac mini M6 32GB | Mac Studio M5 Max 36GB |
|---|---|---|
| Memory | 32 GB unified memory | 36 GB unified memory |
| Usable for a model | ~22.4 GB | ~25.4 GB |
| Memory bandwidth | 170 GB/s | 460 GB/s |
| Chip configuration | 12-core CPU, 12-core GPU | 18-core CPU, 32-core GPU (tied to this memory size) |
| Spec sheet | apple.com | apple.com |
Get an email when a better model fits the Mac Studio M5 Max 36GB
Get one email when a new open-weight model is a better fit for this machine.
The speeds above are estimates. Own a Mac mini M6 32GB? Measure it for real in about a minute and add it to the public leaderboard of measured machines.
Needs Ollama. Sends only the machine type and the measured speed, then removes the 1GB test model.
The speeds above are estimates. Own a Mac Studio M5 Max 36GB? Measure it for real in about a minute and add it to the public leaderboard of measured machines.
Needs Ollama. Sends only the machine type and the measured speed, then removes the 1GB test model.
Speeds are ModelFit engine estimates (est.), not measurements. Context room is memory only: the model's own context window may be smaller.
Prompt reading (est., best case) scales the llama.cpp Apple Silicon table (PP512, 7B Q4_0) to each model's active parameters. That table times a 512-token prompt; attention cost grows with prompt length, so long prompts take longer than shown.
KV-cache figures assume an fp16 cache, the llama.cpp/Ollama default. Standard GQA models use a size-class estimate (8 KV heads x 128 head dim class); hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next) use the exact per-token cost from their published config, since only their sparse full-attention layers cache KV. A q8_0 KV cache roughly halves either figure. Estimates, not measurements.
How this comparison works
Best pick
The highest-ranked local model whose weights plus KV cache still fit the memory budget at the context you choose. At short contexts it is the same pick the homepage wizard gives.
Speed
Estimated tokens per second from the same engine as the rest of the site: chip memory bandwidth from the manufacturer spec sheet and llama.cpp measurements, never above the bandwidth ceiling.
Chip configuration
Apple sells some chips in two configurations. When the memory size decides it (a 36GB M5 Max is always the 32-core GPU version), the tool applies it. When you can choose, a selector appears under the machine.
Prompt reading
How long a Mac takes to read a prompt as long as your context, scaled from the llama.cpp Apple Silicon table. Prompt reading tracks GPU cores, so this is where two configurations of the same chip differ.
Only have a memory size in mind? Use the LLM hardware calculator. Written head-to-heads live on the comparisons page.