Setups with Apple M4 Max: local inference speed (tok/s), sourced
As of 2026-11-14, this dataset has 20 recipes for Apple M4 Max. 14 were measured by the poster. Fastest measured single-device, single-stream text result: 78.7 tok/s with Qwen3.8-Flash-Next on Espetero (Mac Studio M4 Max 128GB; Measured, 2026-10-29; source). 11 entries have a date, from 2026-08-23 to 2026-11-05.
Filter these in the explorer →
| Model | Engine | Hardware | tok/s | Metric | Source level | Date | Source |
|---|---|---|---|---|---|---|---|
| Qwen3.8-Flash-Next | Splash | Apple M4 Max 40C GPU 128GB (model not stated) | 160.5 | decode Splash 160.5 / llama.cpp 42.3 (3.8×); 32K prefill 744.0 / 486.7; 32K cached resend 62.6 / 58.2 ms; 4 concurrent total 144.8 / 89.1; Ollama does not run it | Measured | 2026-10-22 | link |
| Qwen3.8-27B | Splash | Apple M4 Max 128GB (model not stated) | 102.9 | Same file: decode Splash 102.9 / llama.cpp 20.6 (with MTP 28.1) / Ollama 22.2; 32K prefill 332.5 / 159.5 (155.3) / 158.9; 32K cached resend 265.4 / 136.5 (134.3) / 381.7 ms; 4 concurrent total 107.1 / 25.8 (62.9) / 13.0 | Measured | 2026-10-22 | link |
| Qwen3.8-Flash-Next | Espetero | Mac Studio M4 Max 128GB | 78.7 | 1 user oQ4e: 58.6 without MTP, 78.7 with MTP (oMLX 81.1); aggregate (separate) 8 users 170.7 (oMLX 110.0) | Measured | 2026-10-29 | link |
| Qwen3.8-27B | MLX | Mac Studio M4 Max | 65.76 | native stack 65.76 tok/s (MTP depth 3); same GGUF + llama.cpp without MTP: 21.00 | Measured | 2026-09-26 | link |
| Qwen3.8-Flash-Next | oMLX | Mac Studio M4 Max | 63.38 | mean 63.38 tok/s; sequence 68.85, code 52.20, json 69.09 | Measured | 2026-09-26 | link |
| Qwen3.8-27B | Espetero | Mac Studio M4 Max 128GB | 59.4 | 1 user: MLX 4-bit 30.1 without MTP, 59.4 with MTP (mlx-lm 27.5); UD-Q6_K GGUF 20.2, 47.0 with MTP (llama.cpp 19.5, 33.3 with MTP); aggregate (separate) 8 users 161 (mlx-lm 78) | Measured | 2026-10-29 | link |
| Qwen3.8-27B 4-bit | TensorFold | Mac Studio M4 Max | 57.4 | thinking off 76.7 / on 57.4 tok/s (6,071 / 11,895 tokens) | Measured | link | |
| Qwen3.8-27B | oMLX | Mac Studio M4 Max | 53 | 53 tok/s prose, 72 code, draft acceptance rate 78% | Measured | 2026-08-23 | link |
| Qwen3.8-27B-heretic-ara-oQ4e-mtp | mlx-serve | Mac Studio M4 Max | 42 | code 74 / prose 42 tok/s (temp 0, thinking off) | Measured | 2026-10-03 | link |
| DeepSeek-V4-Flash (V4, not 4.1) | ds4 (DwarfStar) | Mac Studio M4 Max 128GB | 30 | ~30 tok/s write; 37K prompt prefill about 2 min | Measured | link | |
| Qwen3.8-Flash-Next | llama.cpp | Mac Studio M4 Max | 23.2 | Reasoning 22.7 / coding 23.0 / strategy 23.2 tok/s | Measured | link | |
| Qwen3.8-27B 4-bit | oMLX | Mac M4 Max | Numeric 154 / 146 / 146; story 51 / 53 / 54; draft off averages 29–31 | Measured | link | ||
| Qwen3.8-Flash-Next | Unspecified | Mac M4 Max / M2 Ultra | 45 / 25 tok/s | Measured | link | ||
| Qwen3.8-27B | MLX | Mac Studio M4 Max | code 74.1; prose 52.8 | Measured | link | ||
| Qwen3.8-Flash-Next | oMLX | Mac M4 Max 128GB | 83 | Best 83 tok/s; 4 streams total 224 | Relayed | link | |
| Qwen3.8-Flash-Next-MLX-Serve-mixed-4 | mlx-serve | Mac Studio/MBP M4 Max 128GB | 60 | 60.0 tok/s decode; 730.0 tok/s prefill | Claimed | 2026-08-27 | link |
| Qwen3.8-27B | TensorFold | Mac Studio M4 Max | 57.2 | DFlash2: code 99.4 / prose 57.2 (24.9 without draft). tok_s uses prose 57.2 | Claimed | 2026-11-05 | link |
| Qwen3.8-27B / Flash-Next / 35B-A3B | MLX | Mac Studio M4 Max 128GB | 27B 70; Flash-Next 83; 35B-A3B 239 tok/s; 27B 4 streams total 119 (old version 72) | Relayed | link | ||
| FastH3 Trim | MLX | Mac M4 Max 36GB | 480p 5s: 925.2 s (V2 not listed) | Official | link | ||
| FastH3 Preview (4 steps) | MLX | Mac M4 Max 36GB | 832×480×124 frames: cold start 465–504 s; INT6 456 s (peak 19.5GiB) | Official | 2026-09-02 | link |
Numbers come from each source. This project did not measure them again. Models and settings differ between rows. Source levels: Official = project or vendor; Measured = the poster ran it (an engine or recipe author must also publish raw data and reproduction steps); Relayed = second-hand; Claimed = prediction or no reproduction details.