Apple M4 Max 47 records: M-series first time in the 200 tok/s club
10-10 batch 47 Apple M4 Max (128GB unified) L2 records avg 196.99 tok/s, approaching 200. M-series first time near desktop N-card. Mac Studio M4 Max is the only 200+ tok/s whole-machine under ¥16000.
Source: araoai
Apple M4 Max 47 records: M-series first time in the 200 tok/s club
10-10 batch has 47 Apple M4 Max (128GB unified) L2 records averaging 196.99 tok/s — M-series first time near desktop N-card.
Overall recap
| Metric | Value |
|---|---|
| Records | 47 |
| Avg decode tok/s | 196.99 |
| Avg VRAM used | 28.33 GB (22% of 128GB) |
| Model range | 0.8B → 27B |
| Best single record | 353.90 tok/s (Qwen3.8-9B + IQ2_XS) |
| Worst single record | 96.32 tok/s (Qwen3.5-27B + Q2_K) |
Key observation: average VRAM 28.33GB = 22% of 128GB — M4 Max 128GB is the only desktop whole-machine that can run 70B+ quantized without offload.
M4 Max vs M3 Max
| Metric | M4 Max (47 records) | M3 Max (34 records) | Gain |
|---|---|---|---|
| Avg tok/s | 196.99 | 152.53 | +29% |
| Avg VRAM used | 28.33 GB | 17.34 GB | +63% |
| Best single | 353.90 | 237.48 | +49% |
M4 → M3 = 29% gain — single generation, close to the 55% RTX 4090 → 5090 step.
5 most dramatic M4 Max records
| Model | Quant | tok/s | VRAM |
|---|---|---|---|
| Qwen3.8-9B | IQ2_XS | 353.90 | 11.17 GB |
| Qwen3.5-0.8B | Q2_K | 352.85 | 3.02 GB |
| Qwen3.5-27B | IQ1_S | 153.29 | 22.80 GB |
| Qwen3.5-4B | Q3_K_M | 296.21 | 7.81 GB |
| Qwen3.5-0.8B | IQ2_XS | 351.07 | 2.91 GB |
Qwen3.8-9B + IQ2_XS + M4 Max = 353 tok/s — 9B model on Apple Silicon hitting 350+ is historic.
M4 Max killer feature: 128GB unified memory
| Setup | VRAM | 27B tok/s | 70B capability |
|---|---|---|---|
| RTX 5090 32GB | 32GB | 175+ | OOM |
| RTX 4090 24GB | 24GB | 91 | OOM |
| Apple M4 Max 128GB | 128GB unified | 153 | 70B IQ2_XS runs |
M4 Max is the only whole-machine that runs 70B quantized — this alone makes the ¥16000 Mac Studio M4 Max the sweet spot for "long-context players".
Build recommendation (Mac fans)
| Budget | Setup | 27B tok/s | 70B capability |
|---|---|---|---|
| ¥16000 | Mac Studio M4 Max 128GB | 153 | ✅ 70B IQ2_XS |
| ¥40000 | Mac Studio M5 Ultra 512GB | est. 280+ | ✅ 200B+ quantized |
| ¥14000 | RTX 5090 + DDR5 64GB | 175 | ❌ OOM |
If you prioritize long-context + 70B+: Mac Studio M4 Max 128GB is the only choice in ¥16000 range. If you prioritize single-card speed: RTX 5090 still owns desktop.
Evidence
- 196.99 average: DB 47 records, hardware_id=26
- 353.90 best: DB record id=1296
- M3 vs M4: DB 34 + 47 records
- "M4 Max is only 70B whole-machine": 128GB unified memory is the only desktop spec
- "M4 Max is 29% above M3 Max": 196.99 / 152.53 = 1.29
Sources
- 47 M4 Max records: DB
benchmark_recordtable, hardware_id=26 - M3 Max 34 records: DB benchmark_record, hardware_id=4
- Best single: DB record id=1296 (Qwen3.8-9B + IQ2_XS)