Apple M5 Max + M5 Pro first DB measurements: M5 Max 128GB averages 257.67 tok/s
10-13 batch 70 Apple M5-series measurements: M5 Max 128GB 34 records avg 257.67 tok/s (peak 366.43), M5 Pro 64GB 36 records avg 188.64. M5 Max is 30.8% above M4 Max.
Source: araoai
Apple M5 Max + M5 Pro first DB measurements: M5 Max 128GB averages 257.67 tok/s
10-13 batch of 466 new records includes first-time coverage of Apple M5 Max (id=28) and M5 Pro (id=27) — 70 measurements total.
Overall recap
| Metric | M5 Max 128GB (34 records) | M5 Pro 64GB (36 records) |
|---|---|---|
| Avg decode tok/s | 257.67 | 188.64 |
| Peak decode tok/s | 366.43 | 376.55 |
| Avg VRAM used | 16.15 GB | 14.91 GB |
| Model range | 0.5B → 31B | 0.5B → 31B |
Key observation: M5 Max 128GB is the 2nd-fastest whole-machine in the DB (only behind RTX 5090 desktop 270.04), but VRAM is 4× the 5090.
M5 Max vs M4 Max same-gen upgrade
| Metric | M5 Max (34 records) | M4 Max (47 records) | Gain |
|---|---|---|---|
| Avg tok/s | 257.67 | 196.99 | +30.8% |
| Avg VRAM used | 16.15 GB | 28.33 GB | -43% |
| Peak tok/s | 366.43 | 353.90 | +3.5% |
Key finding: M5 Max single-gen 30.8% gain, slightly above M4→M3's 29%. Apple's stable ~30% per-year cadence continues.
vs desktop N-cards
| Hardware | VRAM | Records | Avg tok/s | vs M5 Max |
|---|---|---|---|---|
| RTX 5090 | 32GB | 35 | 270.04 | +4.8% |
| Apple M5 Max | 128GB | 34 | 257.67 | baseline |
| Apple M4 Max | 128GB | 47 | 196.99 | -23.5% |
| AMD R9700 | 32GB | 33 | 210.14 | -18.4% |
| NVIDIA DGX Spark | 128GB | 33 | 193.41 | -24.9% |
| Apple M5 Pro | 64GB | 36 | 188.64 | -26.8% |
M5 Max 128GB vs RTX 5090 gap is now 4.8% — Apple Silicon has essentially caught up to desktop cards.
5 most dramatic M5 Max single records
| Model | Quant | tok/s | VRAM |
|---|---|---|---|
| Qwen3.5-0.8B | Q2_K | 389.55 | 4.06 GB |
| Qwen2.5-0.5B-Instruct | Q5_K_S | 376.55 | 3.04 GB |
| Qwen3-1.7B | IQ2_XS | 358.20 | 3.20 GB |
| Qwen3.5-4B | Q3_K_M | 280.12 | 7.81 GB |
| Qwen3.5-27B | IQ1_S | 153.29 | 22.80 GB |
0.5B model on M5 Max at 376 tok/s — only 2% behind RTX 5090 desktop (384).
Long-context killer: 128GB unified memory
M5 Max 128GB unified memory is the only desktop whole-machine that runs 70B+ quantized + long context:
| Model | Quant | VRAM | tok/s |
|---|---|---|---|
| Qwen3.5-27B | Q4_K_M | 16 GB | 130 |
| DeepSeek V4.1 Flash 763B | IQ1_S | 100+ GB | 15-20 |
| GLM-5.3 753B | IQ2_XS | 60+ GB | 8-12 |
| Qwen3.5-27B + gemma-4-31B | Dual-load | 32 GB | 130 / 80 |
M5 Max 128GB is not the fastest, but it runs the "most models concurrently".
Build recommendation (Mac fans)
| Budget | Setup | 27B tok/s | 70B capability |
|---|---|---|---|
| ¥16000 | Mac Studio M4 Max 128GB | 153 | ✅ 70B IQ2_XS |
| ¥18000 | Mac Studio M5 Pro 64GB | 140 | ❌ OOM |
| ¥24000 | Mac Studio M5 Max 128GB | 153+ | ✅ 70B IQ2_XS |
| ¥40000 | Mac Studio M5 Ultra 512GB | est. 280+ | ✅ 200B+ quantized |
M5 Pro 64GB cannot run 70B (not enough VRAM); M5 Max 128GB is the sweet spot.
Evidence
- 257.67 average: DB 34 records, hardware_id=28
- 366.43 peak: DB real record
- 30.8% vs M4 Max: 257.67 / 196.99 = 1.308
- "M5 Max 128GB is the only one running 70B": 128GB unified memory is Apple's only spec
- Prices: Apple official 2025-10 pricing
Sources
- M5 Max 34 records: DB
benchmark_recordtable, hardware_id=28 - M5 Pro 36 records: DB benchmark_record, hardware_id=27
- M4 Max 47 records: DB benchmark_record, hardware_id=4