mlx-serve benchmarks
An inference engine for running open-weight LLMs locally.
15 community reports
mlx-serve results by GPU
Hardware with reported mlx-serve results, with each report's model, quant and speed, newest first. Runs on several cards, with speculative decoding, with batched requests or with part of the model in system RAM say so, since each describes a different setup. Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.
mlx-serve on M4 Max 128GB3 reports
- Qwen3.8 27B43.0 t/s(speculative)
- Qwen3.8 125B · 6B active · mixed 4-8bit60.0 t/s, 730 t/s prompt
- Qwen3.8 125B · 6B active · Q473.8 t/s(speculative)
mlx-serve on M5 Ultra 96GB3 reports
- Qwen3.8 125B · 6B active · iQ-MLX 3.3 bpw1,354 t/s prompt(speculative)
- Qwen3.8 27B · 4.7bpw113.7 t/s, 3,191 t/s prompt(part in system RAM)
- Qwen3.8 125B · 6B active · 4-bit-8bit mix42.7 t/s, 828 t/s prompt(4 requests at once, speculative, part in system RAM)
mlx-serve on M3 Ultra 256GB2 reports
- GLM-5.3 · 4bit48.8 t/s(speculative)
- Qwen3.8 27B · 4-bit84.1 t/s(speculative)
mlx-serve on M5 Max 128GB2 reports
- Qwen3.8 125B · 6B active · Sushi-490.0 t/s, 1,900 t/s prompt(speculative)
- Qwen3.8 125B · 6B active · mixed-4-8bit40.0 t/s(speculative)
mlx-serve on M2 Max 96GB1 report
- Qwen3.8 27B · 4bit20.1 t/s(speculative)
mlx-serve on M3 Max 48GB1 report
- Qwen3.8 27B · 4bit59.6 t/s, 196 t/s prompt(speculative)
mlx-serve on M4 Max 64GB1 report
- Qwen3.8 125B · 6B active · iQ-MLX 3.3bpw52.6 t/s, 762 t/s prompt
mlx-serve on M5 Max 36GB1 report
- Qwen3.6 35B · 3B active · 4-bit178.0 t/s
mlx-serve on M5 Ultra 256GB1 report
- Qwen3.8 125B · 6B active · mixed 4/8-bit154.5 t/s, 4,183 t/s prompt(speculative)
This engine doesn't yet have an editorial profile on llamaperf. The community reports below show how it's been used in practice across different hardware.
Compare GPUs running mlx-serve
Every card is compared on one model, Qwen3.6 35B · 3B active at 4-bit, the size and bit class that most cards here have plain runs of. A card with no run of that model shows none rather than a figure from a different model.
| GPU | VRAM | Reports | Median speed, Qwen3.6 35B · 3B active 4-bit |
|---|---|---|---|
| M4 Max 128GBapple | 128GB | 3 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M5 Max 128GBapple | 128GB | 3 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M5 Ultra 96GBapple | 96GB | 3 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M3 Ultra 256GBapple | 256GB | 2 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M2 Max 96GBapple | 96GB | 1 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M3 Max 48GBapple | 48GB | 1 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M4 Max 64GBapple | 64GB | 1 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M5 Max 36GBapple | 36GB | 1 | Gen 178.0 t/s |
mlx-serve against other engines
Pairs of plain runs on the same card, of the same model size at the same bit class: one device, one request, no speculative decoding, the whole model in memory. Context length and build still differ between the two sides, and each side shows its own.
No matched pair yet. No card has plain runs of one model size at one bit class on mlx-serve and on another engine, so llamaperf can't say how it compares on speed. Add a run.
Top models on mlx-serve
Frequently asked
Is mlx-serve faster than other engines?
llamaperf has no matched comparison for mlx-serve yet: no card has plain runs of the same model size at the same bit class on mlx-serve and on another engine. Speed claims about engines need that pairing, so this page doesn't make one.