Apple · M2 Ultra · Desktop
Best AI Models for Mac Studio M2 Ultra (64 GB)
Runs 27–35B models well and fits 120B MoE; dense 70B fits but is slow at 12 tok/s.
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB | 32.1 tok/s | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB | 32.1 tok/s | |
| 3 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB | 27.5 tok/s | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 91.7 tok/s | |
| 5 | Q4_K_M · 19.3 GB | 77.0 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB | 32.1 tok/s | |
| 2 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB | 27.5 tok/s | |
| 3 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB | 75.9 tok/s | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 91.7 tok/s | |
| 5 | Q4_K_M · 19.3 GB | 77.0 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB | 32.1 tok/s | |
| 2 | Qwen3.5 27BAlibaba Q4_K_M · 17.9 GB | 32.1 tok/s | |
| 3 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB | 27.5 tok/s | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 91.7 tok/s | |
| 5 | Q4_K_M · 19.3 GB | 77.0 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB | 32.1 tok/s | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB | 32.1 tok/s | |
| 3 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB | 75.9 tok/s | |
| 4 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB | 27.5 tok/s | overall |
| 5 | Q4_K_M · 19.3 GB | 77.0 tok/s |
| # | Model | Speed ↓ | Score |
|---|---|---|---|
| 1 | GPT OSS 20BOpenAIHow to run → Q4_K_M · 13.1 GB | 106 tok/s | |
| 2 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 91.7 tok/s | |
| 3 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB | 88.1 tok/s | |
| 4 | Q4_K_M · 19.3 GB | 77.0 tok/s | |
| 5 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB | 75.9 tok/s |
The Mac Studio M2 Ultra is the fastest 64 GB machine: it runs 27–35B models well and fits 120B MoE. macOS keeps about 3.5 GB, leaving 60.5 GB for models. With 43 TFLOPS it reads long prompts faster than most 64 GB machines: a 4K-token prompt takes 26 s. M2 Ultra generation, released 2023; the newer M3 Ultra writes 1% faster with 1.0× the compute.
Mac Studio M2 Ultra strengths and weaknesses
Strengths
- Runs 120B MoE models: GPT OSS 120B at 80.8 tok/s in 59.1 GB
- Room for long context: Qwen3.8 27B at Q4 uses 17.9 GB, leaving 42.6 GB free
Weaknesses
- Slow prompt reading: A 32K-token prompt on Qwen3.8 27B waits 3.5 min for the first token
- No CUDA: vLLM, AWQ, GPTQ and FP8 builds don't run; use GGUF or MLX
Memory for models
60.5 GB of 64 GBEverything on the size ladder fits. Two 27B-class models together use 35.7 GB and leave 24.8 GB free.
Other memory sizes of the same machine first, then machines and GPUs people weigh against it, each with the one reason it is worth a look. Speeds on Qwen3.6 35B A3B.
Output speed depends on memory bandwidth (800 GB/s). Grey rows need more than the 60.5 GB available.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB88.1tok/s
- 14BQwen3 14B · Q4_K_M · 11.9 GB58.9tok/s
- 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB32.1tok/s
- 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB75.9tok/s
- 70BLlama 3.3 70B Instruct · Q4_K_M · 51.3 GB12.0tok/s
- 80.8tok/s
Estimated speeds: Q4_K_M, 4K context, one user. How we estimate
Ultra generations: writing vs compute
M1 Ultra → M3 UltraWriting speed follows memory bandwidth and barely moved: 75.9 to 76.6 tok/s since the M1 Ultra. Compute, which sets how fast prompts are read, grew 1.4×. Indexed to the M1 Ultra; speeds on Qwen3.6 35B A3B, same 64 GB.
All 1168 models on the Mac Studio M2 Ultra
Featured onlyBest quant that fits in 60.5 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works
| Model | Quant | VRAM | Speed | Context | Score | Fit |
|---|---|---|---|---|---|---|
DeepSeek v2 LiteChat | Q4_K_M | 10.2 GB | 117.9 t/s | 164K | S94, Runs great | |
| Q4_K_M | 2.3 GB | 643.7 t/s | 33K | S100, Runs great | ||
OLMoE 1B 7B 0924Chat | Q4_K_M | 5.0 GB | 210.0 t/s | 4K | S100, Runs great | |
MythoMax L2 13BChat | Q4_K_M | 11.5 GB | 57.3 t/s | 4K | A82, Runs well | |
| Q4_K_M | 6.3 GB | 111.1 t/s | 262K | S93, Runs great | ||
MiniCPM V 2 6Vision | Q4_K_M | 7.9 GB | 104.7 t/s | — | S92, Runs great | |
| Q4_K_M | 20.7 GB | 27.8 t/s | 4K | A68, Runs well | ||
| BF16 | 24.8 GB | 69.2 t/s | 131K | S86, Runs great | ||
Kumru 2B BaseChat | Q4_K_M | 2.2 GB | 304.3 t/s | 8K | S100, Runs great | |
Llama 3.1 70B InstructChatFunctions | Q4_K_M | 51.3 GB | 12.0 t/s | 131K | C44, Usable | |
| Q4_K_M | 12.1 GB | 126.3 t/s | 33K | S95, Runs great | ||
| BF16 | 20.4 GB | 31.4 t/s | — | A71, Runs well | ||
Qwen3 0.6B BaseChat | Q4_K_M | 1.6 GB | 717.9 t/s | 33K | S100, Runs great | |
Qwen3 4B BaseChat | Q4_K_M | 4.2 GB | 193.1 t/s | 33K | S99, Runs great | |
Deepseek Coder 33B InstructChatCode | Q4_K_M | 24.5 GB | 26.9 t/s | 16K | A68, Runs well |
Specs
- Chip
- M2 Ultra
- CPU cores
- 24
- GPU cores
- 60
- Unified memory
- 64 GB
- For models
- 60.5 GB (macOS keeps 3.5 GB)
- Memory bandwidth
- 800 GB/s
- FP16 (dense)
- 43.0 TFLOPS
- Neural Engine
- 32 TOPS
- Released
- 2023-06-13
Mac Studio M2 Ultra FAQ
How much memory is available for AI on the Mac Studio M2 Ultra (64 GB)?
About 60.5 GB of the 64 GB. macOS keeps roughly 3.5 GB for itself and other apps.
What is the largest model the Mac Studio M2 Ultra (64 GB) can run?
GPT OSS 120B (116.8B) at Q4_0 uses 59.1 GB, 98% of the available memory, and runs at about 80.8 tok/s.
How long does the Mac Studio M2 Ultra (64 GB) take to read a long prompt?
On Qwen3.8 27B, about 26 s for a 4K-token prompt and 3.5 min for a 32K-token document before the first word appears.
Is renting a GPU cheaper than buying the Mac Studio M2 Ultra?
An A100 (80 GB) rents from $1.07/hr.