News
Apple M1/M2 revival: old Macs can run LLM — M1 avg 90 tok/s, M2 Ultra 192GB runs 70B
10-16 batch 209 Apple M1-M2 measurements first time in DB: M1 90.65, M1 Pro 118.71, M1 Max 147.03, M2 Pro 123.14, M2 Ultra 172.06 tok/s. **M2 Ultra 192GB unified is the only ¥24000 whole-machine that runs 70B**.
Source: araoai
Apple M1/M2 revival: old Macs can run LLM — M1 avg 90 tok/s, M2 Ultra 192GB runs 70B
10-16 batch has 209 Apple M1/M2-series L2 records — Apple 2020-2023 old Mac ecosystem first time in DB.
Overall recap (5 M-series chips)
| Hardware | VRAM | Records | Avg tok/s | Peak | Form factor |
|---|---|---|---|---|---|
| M2 Ultra | 192GB | 47 | 172.06 | 393.85 | Mac Studio / Mac Pro |
| M1 Max | 64GB | 41 | 147.03 | 395.33 | MBP 16 / Mac Studio |
| M2 Pro | 32GB | 41 | 123.14 | 271.73 | MBP 14 / Mac mini |
| M1 Pro | 32GB | 41 | 118.71 | 263.91 | MBP 14/16 |
| M1 | 16GB | 39 | 90.65 | 231.36 | MBP 13 / Mac mini / iPad |
| For ref: M5 Max (34 records) | 128GB | — | 257.67 | 366.43 | Mac Studio |
| For ref: M4 Max (47 records) | 128GB | — | 196.99 | 353.90 | Mac Studio |
Key observations:
- M1 hits 90 tok/s — 2020 MBP 13 + M1 + 16GB unified runs Qwen3.5-27B IQ2_XS, enough for daily use
- M2 Ultra 192GB unified is the only ¥24000 whole-machine that runs 70B quantized (vs Mac Studio M5 Max 128GB ¥24000 + DGX Spark 128GB ¥30000)
- M1 Max peak 395.33 tok/s — Apple Silicon old flagship still in 200+ tok/s club
M2 Ultra 192GB killer feature
| Model | Quant | VRAM used | tok/s |
|---|---|---|---|
| Qwen3.5-27B | Q4_K_M | 16 GB | 130 |
| Qwen3.5-27B | IQ1_S | 13 GB | 175 |
| DeepSeek V4.1 Flash 763B | IQ1_S | 150+ GB | 3-4 |
| GLM-5.3 753B | IQ2_XS | 130+ GB | 5-7 |
| Qwen3.5-27B + gemma-4-31B | Dual-load | 32 GB | 130 / 80 |
M2 Ultra 192GB unique point: fits V4.1 Flash 763B IQ1_S (150GB+) + run macOS Sonoma simultaneously, with 40GB+ headroom. M5 Max 128GB can't fit 700B+.
Old Mac revival checklist
| Your old Mac | Fit | Recommended model | tok/s |
|---|---|---|---|
| MBP 13 M1 16GB | ✅ strong | Qwen3.5-4B IQ4_XS | 90 |
| MBP 14 M1 Pro 32GB | ✅ strong | Qwen3.5-9B Q4_K_M | 118 |
| MBP 16 M1 Max 64GB | ✅ very strong | Qwen3.5-27B IQ2_XS | 147 |
| Mac mini M2 Pro 32GB | ✅ strong | Qwen3.5-9B Q4_K_M | 123 |
| Mac Studio M2 Ultra 192GB | ⭐ top | DeepSeek V4.1 Flash 763B | 3.5 |
Build recommendation: don't upgrade the Mac, 4-year-old M1 Max 64GB + Qwen3.5-27B = 147 tok/s is enough for local inference. M1-M2 isn't e-waste, it's the LLM inference sweet spot.
Evidence
- 90.65/118.71/147.03/123.14/172.06 averages: DB 209 records
- M1 Max peak 395.33: DB record id=2044
- M2 Ultra 192GB: Apple official spec
- "M1 hits 90": DB 39 records
- "M2 Ultra is the only ¥24000 70B": based on current pricing + VRAM comparison
Sources
- 209 M1-M2 records: DB
benchmark_recordtable, hardware_id IN (99, 100, 101, 102, 103) - M5 Max 34 records: DB benchmark_record, hardware_id=28
- M4 Max 47 records: DB benchmark_record, hardware_id=4