Twoody
Local LLM calculator
The best local LLM for your Mac Studio
From the M1 Max to the M5 Ultra with 512 GB: the largest models a Mac can hold, and how fast they write.
| Your computer | Memory | Recommended — runs comfortably | Writes | One click in Twoody |
|---|---|---|---|---|
M5 Max — MacBook Pro, Mac Studio |
36 GB | GLM-4.7-Flash | 100+ tokens/s | Qwen3.5 9B |
| 48 GB | GLM-4.7-Flash | 100+ tokens/s | Qwen3.5 9B | |
| 64 GB | GLM-4.7-Flash | 100+ tokens/s | Qwen3.5 9B | |
| 128 GB | Qwen3-Next 80B-A3B | 96–193 tokens/s | Qwen3.5 9B | |
M5 Ultra — Mac Studio |
96 GB | gpt-oss-120b | 92–163 tokens/s | Qwen3.5 9B |
| 256 GB | Qwen3 235B-A22B | 28–50 tokens/s | Qwen3.5 9B | |
| 512 GB | Qwen3 235B-A22B | 28–50 tokens/s | Qwen3.5 9B | |
M3 Ultra — Mac Studio |
96 GB | gpt-oss-120b | 63–90 tokens/s | Qwen3.5 9B |
| 256 GB | Qwen3 235B-A22B | 20–28 tokens/s | Qwen3.5 9B | |
| 512 GB | Qwen3 235B-A22B | 20–28 tokens/s | Qwen3.5 9B | |
M4 Max — MacBook Pro, Mac Studio |
36 GB | GLM-4.7-Flash | 73–106 tokens/s | Qwen3.5 9B |
| 48 GB | gpt-oss-20b | 85–122 tokens/s | Qwen3.5 9B | |
| 64 GB | gpt-oss-20b | 85–122 tokens/s | Qwen3.5 9B | |
| 128 GB | Qwen3-Next 80B-A3B | 53–107 tokens/s | Qwen3.5 9B | |
M2 Ultra — Mac Studio |
64 GB | gpt-oss-20b | 89–128 tokens/s | Qwen3.5 9B |
| 128 GB | gpt-oss-120b | 63–91 tokens/s | Qwen3.5 9B | |
| 192 GB | Qwen3 235B-A22B | 20–28 tokens/s | Qwen3.5 9B | |
M2 Max — MacBook Pro, Mac Studio |
32 GB | gpt-oss-20b | 62–90 tokens/s | Qwen3.5 9B |
| 64 GB | gpt-oss-20b | 62–90 tokens/s | Qwen3.5 9B | |
| 96 GB | Qwen3-Next 80B-A3B | 39–78 tokens/s | Qwen3.5 9B | |
M1 Ultra — Mac Studio |
64 GB | gpt-oss-20b | 75–107 tokens/s | Qwen3.5 9B |
| 128 GB | gpt-oss-120b | 52–75 tokens/s | Qwen3.5 9B | |
M1 Max — MacBook Pro, Mac Studio |
32 GB | gpt-oss-20b | 56–80 tokens/s | Qwen3.5 9B |
| 64 GB | gpt-oss-20b | 56–80 tokens/s | Qwen3.5 9B |
Estimates, not measurements, computed from public llama.cpp benchmarks. Figures reviewed on September 26, 2026.
With 128 to 512 GB of unified memory, a Mac Studio holds models no graphics card can: gpt-oss-120b, Qwen3.5 122B-A10B, even Qwen3 235B-A22B.
Its Max and Ultra chips read their memory at 400 GB/s to 1.2 TB/s: mixtures of experts answer at dozens of tokens per second.
Quiet even under load, it keeps its speed through long sessions.
Questions about running an LLM locally
Which LLM for a Mac Studio with an Ultra chip?
With 256 or 512 GB, Qwen3 235B-A22B, the largest model here, runs comfortably; gpt-oss-120b and Qwen3.5 122B-A10B write faster.
Is 128 GB enough for a 70B model?
Yes, comfortably: Llama 3.3 70B needs about 48 GB at 4-bit. With 128 GB, mixtures of experts such as gpt-oss-120b fit too.
Max or Ultra for a local LLM?
The Ultra doubles the memory bandwidth and the memory: faster answers with large models, and the largest models. For models up to 70B, a Max is enough.
The best local LLM, by computer
Twoody is in private beta.
On Mac, Windows and Linux, free, with no account and no subscription — and on iPhone and Android, with Twoody on your computer. Leave your email: we will write to you when Twoody opens to you.