Compare two machines for local AI
Configure two Macs or graphics cards side by side and see what each one runs, how fast, and what it costs, from the same engine as the Mac configurator.
Compare configurator
Machine.
Which B?Chip.
Bandwidth sets generation speed.Generation speed follows memory bandwidth in this estimate; GPU cores speed up prompt reading.
Processing power.
What the extra cores change.The 40-core GPU ships with 48 / 64 / 128 GB, the 32-core with 36 GB: generation about 27% faster (460 GB/s vs 614 GB/s), prompt reading about 25% faster (ModelFit estimate, partly scaled by GPU core count).
Speeds for Qwen3.6 35B-A3B, est. Prompt reading is scaled from llama.cpp's 7B Q4_0 prompt speed (PP512): source.
Memory.
Sets which models fit.Use.
What will you run? Applies to both.Mac mini M6 32GB vs Mac Studio M5 Max 36GB
| Model | Quant | Run | ||||||
|---|---|---|---|---|---|---|---|---|
| 1 | 26B moe 4B | Q4_K_M | 16.0 GB | 29.2 | Runs well 38 | Runs well 84 | ||
| 2 | 35B moe 3B | Q4_K_M | 22.0 GB | 32.0 | Runs well 59 | Runs well 128 | ||
| 3 | 24B moe 2B | Q4_K_M | 14.0 GB | n/a | Runs well 66 | Runs well 141 | ||
| 4 | 35B | Q4_K_M | 21.2 GB | n/a | Runs well 61 | Runs well 131 | ||
| 5 | 30B moe 3B | Q4_K_M | 19.0 GB | 23.0 | Runs well 40 | Runs well 90 | ||
| 6 | 27B | Q4_K_M | 16.5 GB | 44.4 | Runs well 10 | Runs well 23 | ||
| 7 | 33B moe 3B | Q4_K_M | 20.3 GB | 23.9 | Runs well 55 | Runs well 120 | ||
| 8 | 30B moe 3B | IQ4_XS | 18.9 GB | 21.8 | Runs well 47 | Runs well 103 | no Ollama tag | |
| 9 | 35B moe 3B | Q4_K_M | 20.0 GB | 30.5 | Runs well 61 | Runs well 131 | ||
| 10 | 31B | Q4_K_M | 20.0 GB | 32.7 | Runs well 8 | Runs well 18 | ||
| 11 | 21B moe 3.6B | MXFP4 | 13.8 GB | 17.5 | Runs well 38 | Runs well 85 | ||
| 12 | 30B moe 3B | Q4_K_M | 19.0 GB | 21.5 | Runs well 41 | Runs well 90 | ||
| 13 | 9B | Q4_K_M | 7.0 GB | 24.2 | Lots of room 31 | Lots of room 69 | ||
| 14 | 12B | Q8_0 | 12.8 GB | 21.6 | Runs well 12 | Runs well 28 | ||
| 15 | 24B | Q4_K_M | 15.2 GB | n/a | Runs well 10 | Runs well 24 | ||
| 16 | 27B | Q4_K_M | 18.0 GB | 35.2 | Runs well 10 | Runs well 23 | ||
| 17 | 29.3B | Q4_K_M | 16.5 GB | 23.3 | Runs well 9 | Runs well 20 | ||
| 18 | 27B | Q4_K_M | 16.0 GB | 33.5 | Runs well 10 | Runs well 23 | ||
| 19 | 4.5B | Q8_0 | 7.5 GB | 13.3 | Lots of room 31 | Lots of room 69 | ||
| 20 | 30B moe 3B | Q6_K | 24.0 GB | n/a | Tight 31 | Runs well 83 | ||
| 21 | 4B | Q4_K_M | 3.5 GB | 19.0 | Lots of room 54 | Lots of room 117 | ||
| 22 | 13.5B | Q4_K_M | 9.1 GB | 9.8 | Runs well 17 | Lots of room 39 |
Click a row to map its memory on both machines in Fig. 1Quality: LLM Stats Score · Data by LLM StatsOllama commands are registry-verified
Ranked #49 on LLM Stats; today's leader, Claude Opus 5.5, scores 60.4.
The best-scored model this configuration runs.
Data by LLM Stats · 2026-10-22 · Full benchmarks on llm-stats.com
Ranked #49 on LLM Stats; today's leader, Claude Opus 5.5, scores 60.4.
The best-scored model this configuration runs.
Data by LLM Stats · 2026-10-22 · Full benchmarks on llm-stats.com
How this comparison works
Top pick
The highest-ranked local model that runs comfortably in the memory budget: the same pick the homepage wizard and the Mac configurator give. Click any row of Table 1 to compare that model on both machines.
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. The Processing power step shows both, with what the extra cores change. When the memory size decides it (a 36GB M5 Max is always the 32-core GPU version), only that one is offered.
Prompt reading
How fast a Mac reads a prompt, 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.