M5 (32GB) holds 279 of the models in our catalogue and is, in practice, a Q4_K_S card — the largest it takes is ERNIE 4.5 VL 28B A3B Thinking at Q4_K_S.
Neighbours in memory rather than in price: memory decides whether a card can do the job at all, so two cards of the same size at different prices is the comparison you are making. Ordered by memory, then bandwidth — not by the value column, which is worked out from bandwidth and price alone and therefore rewards a cheap card whatever its software stack does to that bandwidth in practice. Read it as one input, not as a ranking.
Apple M5 (32GB) — 32 GB unified, 21.3 GB usable.
- BRAND
- Apple
- UNIFIED MEMORY
- 32 GB
- USABLE BY A MODEL
- 21.3 GB
- BANDWIDTH
- 120 GB/s
- FP16 COMPUTE
- 5.5 TFLOPS
- TDP
- 15 W
- ARCHITECTURE
- M5
With 21.3 GB of its 32 GB reaching a model and 120 GB/s bandwidth, this machine handles models up to 27.2B parameters.
Speed ≈ bandwidth / model_size × efficiency. A 7B model at Q4 runs at ~14 tok/s.
| MODEL | SIZE | VRAM Q4 | TOK/S | AVG |
|---|---|---|---|---|
| gemma-2-27b | 27.2B | 17.1 GB | 4 | 34.6 |
| Gemma 4 26B A4B | 26B | 16.4 GB | 27 | 47.9 |
| Aria 25B A3.9B | 25.3B | 16.0 GB | 27 | 64.8 |
| Mistral-Small-24B | 24B | 15.2 GB | 4 | 25.0 |
| Mistral-Small-3.1-24B | 24B | 15.2 GB | 4 | 28.8 |
| Magistral Small 24B | 24B | 15.2 GB | 4 | 47.0 |
| Devstral Small 2 24B | 24B | 15.2 GB | 4 | 33.4 |
| Mistral-Small-3.2-24B | 24B | 15.2 GB | 4 | 44.4 |
| LFM2 24B A2B | 24B | 15.2 GB | 53 | 19.1 |
| Devstral Small 22B | 23.57B | 14.9 GB | 5 | 35.5 |
| Codestral 22B | 22.2B | 14.1 GB | 5 | 50.1 |
| Mistral Small 22B | 22.2B | 14.1 GB | 5 | 35.2 |
| SOLAR-Pro 22B | 22.1B | 14.0 GB | 5 | 44.2 |
| ERNIE 4.5 21B A3B | 21.95B | 13.9 GB | 36 | — |
| GPT-OSS 20B | 21B | 13.3 GB | 30 | 52.9 |
| Reka Flash 3 | 21B | 13.3 GB | 5 | 36.2 |
| Reka Flash 3.1 | 21B | 13.3 GB | 5 | 33.4 |
| InternLM2 20B | 19.8B | 12.6 GB | 5 | 45.1 |
| InternLM2.5 20B | 19.8B | 12.6 GB | 5 | 50.9 |
| Ling-lite 16.8B | 16.8B | 10.8 GB | 44 | — |