Apple · M5 Max · Desktop
Best AI Models for Mac Studio M5 Max (36 GB)
Runs 27–35B models well; 70B doesn't fit in 32.5 GB.
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB | 18.5 tok/s | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB | 18.5 tok/s | |
| 3 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB | 15.8 tok/s | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 68.8 tok/s | |
| 5 | Q4_K_M · 19.3 GB | 67.2 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB | 18.5 tok/s | |
| 2 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB | 15.8 tok/s | |
| 3 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB | 58.8 tok/s | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 68.8 tok/s | |
| 5 | Q4_K_M · 19.3 GB | 67.2 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB | 18.5 tok/s | |
| 2 | Qwen3.5 27BAlibaba Q4_K_M · 17.9 GB | 18.5 tok/s | |
| 3 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB | 15.8 tok/s | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 68.8 tok/s | |
| 5 | Q4_K_M · 19.3 GB | 67.2 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB | 18.5 tok/s | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB | 18.5 tok/s | |
| 3 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB | 58.8 tok/s | |
| 4 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB | 15.8 tok/s | overall |
| 5 | Q4_K_M · 19.3 GB | 67.2 tok/s |
| # | Model | Speed ↓ | Score |
|---|---|---|---|
| 1 | GPT OSS 20BOpenAIHow to run → Q4_K_M · 13.1 GB | 77.9 tok/s | |
| 2 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 68.8 tok/s | |
| 3 | Q4_K_M · 19.3 GB | 67.2 tok/s | |
| 4 | Qwen3 30B A3B Thinking 2507Alibaba Q4_K_M · 19.4 GB | 62.6 tok/s | |
| 5 | Qwen3 30B A3B Instruct 2507Alibaba Q4_K_M · 19.4 GB | 62.6 tok/s |
The Mac Studio M5 Max is the fastest 36 GB machine: it runs 27–35B models well. macOS keeps about 3.5 GB, leaving 32.5 GB for models. With 54 TFLOPS it reads long prompts faster than most 36 GB machines: a 4K-token prompt takes 21 s. M5 Max generation, released 2026.
Mac Studio M5 Max strengths and weaknesses
Strengths
- Runs 35B MoE models: Qwen3.6 35B A3B at 58.8 tok/s in 22 GB
- Room for long context: Qwen3.8 27B at Q4 uses 17.9 GB, leaving 14.6 GB free
Weaknesses
- 70B models don't fit: Llama 3.3 70B Instruct needs 51.3 GB at Q4 vs 32.5 GB usable here
- Slow prompt reading: A 32K-token prompt on Qwen3.8 27B waits 2.8 min for the first token
- No CUDA: vLLM, AWQ, GPTQ and FP8 builds don't run; use GGUF or MLX
Memory for models
32.5 GB of 36 GB70B models need 51.3 GB at Q4, more than the 32.5 GB available. The 48 GB config fits Mixtral 8x7B Instruct v0.1 →
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 (460 GB/s). Grey rows need more than the 32.5 GB available.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB50.6tok/s
- 14BQwen3 14B · Q4_K_M · 11.9 GB33.9tok/s
- 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB18.5tok/s
- 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB58.8tok/s
- 70BLlama 3.3 70B Instruct · does not fitneeds 51.3 GB
- 120B MoEGPT OSS 120B · does not fitneeds 70.8 GB
Estimated speeds: Q4_K_M, 4K context, one user. How we estimate
Max generations: writing vs compute
M1 Max → M5 MaxWriting speed follows memory bandwidth and barely moved: 54.5 to 58.8 tok/s since the M1 Max. Compute, which sets how fast prompts are read, grew 3.5×. Indexed to the M1 Max; speeds on Qwen3.6 35B A3B, same 32 GB.
All 1115 models on the Mac Studio M5 Max
Featured onlyBest quant that fits in 32.5 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works
| Model | Quant | VRAM | Speed | Context | Score | Fit |
|---|---|---|---|---|---|---|
Qwen2.5 0.5BChatFunctions | Q4_K_M | 0.8 GB | 519.4 t/s | 33K | S100, Runs great | |
C4ai Command R V01ChatFunctions | Q4_K_M | 27.0 GB | 13.9 t/s | — | B49, Decent | |
| BF16 | 1.6 GB | 527.9 t/s | — | S100, Runs great | ||
Starcoder2 15BChatCode | Q4_K_M | 11.2 GB | 32.1 t/s | 16K | A71, Runs well | |
| Q4_K_M | 30.4 GB | 12.2 t/s | — | C30, Usable | ||
SmolLM2 360MChat | Q4_K_M | 0.8 GB | 536.7 t/s | 8K | S100, Runs great | |
Deepseek Coder 7B Instruct V1.5ChatCode | Q4_K_M | 6.5 GB | 59.1 t/s | 4K | A83, Runs well | |
Yi 6BChat | Q4_K_M | 4.2 GB | 79.1 t/s | 4K | S87, Runs great | |
| Q4_K_M | 17.8 GB | 21.7 t/s | 131K | B62, Decent | ||
Qwen2.5 Math 7B InstructChatMathFunctions | Q4_K_M | 5.1 GB | 64.5 t/s | 4K | A84, Runs well | |
Llama 3.2 1BChat | Q4_K_M | 2.3 GB | 392.7 t/s | — | S100, Runs great | |
SmolLM 360MChat | Q4_K_M | 0.6 GB | 536.7 t/s | 2K | S100, Runs great | |
Qwen2.5 1.5BChatFunctions | Q4_K_M | 1.7 GB | 251.6 t/s | 131K | S100, Runs great | |
Falcon 7BChat | Q4_K_M | 7.3 GB | 67.6 t/s | — | S85, Runs great | |
Meta Llama 3 8BChat | Q4_K_M | 6.2 GB | 59.7 t/s | 8K | A84, Runs well |
Specs
- Chip
- M5 Max
- CPU cores
- 18
- GPU cores
- 32
- Unified memory
- 36 GB
- For models
- 32.5 GB (macOS keeps 3.5 GB)
- Memory bandwidth
- 460 GB/s
- FP16 (dense)
- 54.4 TFLOPS
- Neural Engine
- 38 TOPS
- Released
- 2026-09-22
Mac Studio M5 Max FAQ
How much memory is available for AI on the Mac Studio M5 Max (36 GB)?
About 32.5 GB of the 36 GB. macOS keeps roughly 3.5 GB for itself and other apps.
What is the largest model the Mac Studio M5 Max (36 GB) can run?
Qwen3.6 35B A3B (36.0B) at Q4_K_M uses 22.0 GB, 68% of the available memory, and runs at about 58.8 tok/s.
Can the Mac Studio M5 Max (36 GB) run 70B models?
No. Llama 3.3 70B Instruct needs 51.3 GB at Q4, more than the 32.5 GB available.
How long does the Mac Studio M5 Max (36 GB) take to read a long prompt?
On Qwen3.8 27B, about 21 s for a 4K-token prompt and 2.8 min for a 32K-token document before the first word appears.
Is renting a GPU cheaper than buying the Mac Studio M5 Max?
An A40 (48 GB) rents from $0.35/hr.