Which local LLMs can the Apple M5 Pro (Mac mini, MacBook Pro) run?
Unified memory — by default the GPU can use about 70% of it
Assumes 8k context and an f16 KV cache. Verdicts are shown for each memory size.
Specs
- Memory options
- 24 GB · 48 GB · 64 GB
- Memory bandwidth
- 307.0 GB/s
- FP16 compute
- 8.3 TFLOPS
- Launch year
- 2026
- Macs with this chip
- Mac mini: 24 GB · 48 GB · 64 GB
- MacBook Pro: 24 GB · 48 GB · 64 GB
- Also written as
- M5 Pro
Verdicts at a glance
At Q4_K_M (or the closest available quant) with 8k context.
- 24 GBRuns great26 modelsRuns well4 modelsRuns slowly10 modelsWon't run19 models8 more models run at a lower quant.
- 48 GBRuns great33 modelsRuns well10 modelsRuns slowly6 modelsWon't run10 models4 more models run at a lower quant.
- 64 GBRuns great33 modelsRuns well10 modelsRuns slowly8 modelsWon't run8 models2 more models run at a lower quant.
How to read the verdicts
- Runs great
- Fully on the GPU at 20 tok/s or more
- Runs well
- Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
- Runs slowly
- 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
- Won't run
- Does not fit, or under 2 tok/s
Every model on the Apple M5 Pro
Scroll sideways to see every column.
| Model | Verdict | Speed | Quant | Memory | Runs as | Context | Notes |
|---|---|---|---|---|---|---|---|
| MiniCPM5 1B | Runs great | est. 261.8 tok/s209.4–314.1calibrated estimate ±20% | Q4_K_M | 1.5 / 44.8 GB | Unified memory | up to 128k |
|
| Llama 3.2 1B | Runs great | est. 216.4 tok/s173.1–259.6calibrated estimate ±20% | Q4_K_M | 1.7 / 44.8 GB | Unified memory | up to 128k | — |
| Ling-3.0 Tiny | Runs great | est. 189.8 tok/s132.9–246.8theoretical estimate ±30% | Q4_K_M | 5.5 / 44.8 GB | Unified memory | up to 128k |
|
| EXAONE 4.0 1.2B | Runs great | est. 177.7 tok/s142.1–213.2calibrated estimate ±20% | Q4_K_M | 1.9 / 44.8 GB | Unified memory | up to 64k |
|
| LFM2.5 2.6B | Runs great | est. 129.3 tok/s90.5–168.1theoretical estimate ±30% | Q4_K_M | 2.4 / 44.8 GB | Unified memory | up to 128k |
|
| HyperCLOVA X SEED 1.5B | Runs great | est. 128.5 tok/s102.8–154.2calibrated estimate ±20% | Q4_K_M | 2.4 / 44.8 GB | Unified memory | up to 16k | — |
| MiniCPM5 2B | Runs great | est. 122.0 tok/s97.6–146.4calibrated estimate ±20% | Q4_K_M | 2.5 / 44.8 GB | Unified memory | up to 128k |
|
| Gemma 4 E2B | Runs great | est. 119.1 tok/s83.4–154.9theoretical estimate ±30% | Q4_K_M | 3.7 / 44.8 GB | Unified memory | up to 128k |
|
| Mellum2.1 12B-A2.5B Thinking | Runs great | est. 111.4 tok/s89.1–133.7calibrated estimate ±20% | Q4_K_M | 8.8 / 44.8 GB | Unified memory | up to 128k |
|
| MiniCPM-V 4.7 35B-A3B | Runs great | est. 81.0 tok/s64.8–97.3calibrated estimate ±20% | Q4_K_M | 21.9 / 44.8 GB | Unified memory | up to 128k |
|
| Spark-X2.5 4B | Runs great | est. 80.4 tok/s64.3–96.5calibrated estimate ±20% | Q4_K_M | 3.5 / 44.8 GB | Unified memory | up to 128k |
|
| Llama 3.2 3B | Runs great | est. 78.8 tok/s63.1–94.6calibrated estimate ±20% | Q4_K_M | 3.5 / 44.8 GB | Unified memory | up to 64k | — |
| gpt-oss-20b | Runs great | est. 77.9 tok/s62.4–93.5calibrated estimate ±20% | MXFP4 | 12.9 / 44.8 GB | Unified memory | up to 128k |
|
| Nemotron 3.5 Lightning 30B-A3B | Runs great | est. 75.6 tok/s52.9–98.3theoretical estimate ±30% | Q4_K_M | 26.1 / 44.8 GB | Unified memory | up to 256k |
|
| Qwen3.5 4B | Runs great | est. 75.3 tok/s60.2–90.4calibrated estimate ±20% | Q4_K_M | 3.7 / 44.8 GB | Unified memory | up to 128k |
|
| Qwen3.6 35B-A3B | Runs great | est. 75.3 tok/s60.2–90.4calibrated estimate ±20% | Q4_K_M | 23.0 / 44.8 GB | Unified memory | up to 128k |
|
| Kanana 1.5 15.7B-A3B | Runs great | est. 66.9 tok/s53.5–80.3calibrated estimate ±20% | Q4_K_M | 12.1 / 44.8 GB | Unified memory | up to 32k | — |
| Qwen3.5 35B-A3B | Runs great | est. 66.3 tok/s53.0–79.5calibrated estimate ±20% | Q4_K_M | 23.4 / 44.8 GB | Unified memory | up to 128k |
|
| Gemma 4 E4B | Runs great | est. 65.1 tok/s45.6–84.7theoretical estimate ±30% | Q4_K_M | 5.7 / 44.8 GB | Unified memory | up to 128k |
|
| Qwen3 30B-A3B (2507) | Runs great | est. 64.8 tok/s51.8–77.8calibrated estimate ±20% | Q4_K_M | 19.9 / 44.8 GB | Unified memory | up to 64k | — |
| Gemma 4 26B-A4B | Runs great | est. 55.6 tok/s44.5–66.8calibrated estimate ±20% | Q4_K_M | 17.7 / 44.8 GB | Unified memory | up to 128k |
|
| Qwen3 8B | Runs great | 45.4 tok/s8k estimate 37.4 tok/s (29.9–44.9, calibrated estimate ±20%)measured (1 run, 512 context) | Q4_K_M | 6.8 / 44.8 GB | Unified memory | up to 16k |
|
| K2-Horizon MoVA 36B-A4B | Runs great | est. 40.7 tok/s28.5–52.9theoretical estimate ±30% | Q4_K_M | 24.6 / 44.8 GB | Unified memory | up to 16k |
|
| DeepSeek R1 Distill Llama 8B | Runs great | est. 38.9 tok/s31.1–46.7calibrated estimate ±20% | Q4_K_M | 6.6 / 44.8 GB | Unified memory | up to 16k |
|
| Kanana 1.5 8B | Runs great | est. 38.9 tok/s31.1–46.7calibrated estimate ±20% | Q4_K_M | 6.6 / 44.8 GB | Unified memory | up to 32k | — |
| Llama 3.1 8B | Runs great | est. 38.9 tok/s31.1–46.7calibrated estimate ±20% | Q4_K_M | 6.6 / 44.8 GB | Unified memory | up to 32k | — |
| Nanbeige4.2 3B | Runs great | est. 38.1 tok/s26.7–49.5theoretical estimate ±30% | Q4_K_M | 4.7 / 44.8 GB | Unified memory | up to 16k |
|
| Qwen3.5 9B | Runs great | est. 38.0 tok/s30.4–45.6calibrated estimate ±20% | Q4_K_M | 6.7 / 44.8 GB | Unified memory | up to 32k |
|
| DeepSeek R1 0528 Qwen3 8B | Runs great | est. 37.4 tok/s29.9–44.9calibrated estimate ±20% | Q4_K_M | 6.8 / 44.8 GB | Unified memory | up to 16k |
|
| K2-Horizon 7B | Runs great | est. 34.3 tok/s27.5–41.2calibrated estimate ±20% | Q4_K_M | 7.4 / 44.8 GB | Unified memory | up to 8k |
|
| Gemma 4 12B | Runs great | est. 32.2 tok/s25.7–38.6calibrated estimate ±20% | Q4_K_M | 7.8 / 44.8 GB | Unified memory | up to 32k |
|
| Gemma 3 12B | Runs great | est. 29.8 tok/s23.8–35.7calibrated estimate ±20% | Q4_K_M | 8.4 / 44.8 GB | Unified memory | up to 64k | — |
| Phi-4 | Runs great | est. 21.7 tok/s17.4–26.1calibrated estimate ±20% | Q4_K_M | 11.3 / 44.8 GB | Unified memory | up to 8k | — |
| HyperCLOVA X SEED Think 14B | Runs well | est. 22.9 tok/s16.0–29.8theoretical estimate ±30% | Q4_K_M | 10.8 / 44.8 GB | Unified memory | up to 64k |
|
| Qwen3 14B | Runs well | est. 22.6 tok/s18.0–27.1calibrated estimate ±20% | Q4_K_M | 10.9 / 44.8 GB | Unified memory | up to 32k |
|
| DeepSeek R1 Distill Qwen 14B | Runs well | est. 22.0 tok/s17.6–26.4calibrated estimate ±20% | Q4_K_M | 11.2 / 44.8 GB | Unified memory | up to 32k |
|
| Mistral Small 3.2 24B | Runs well | est. 14.9 tok/s11.9–17.9calibrated estimate ±20% | Q4_K_M | 16.2 / 44.8 GB | Unified memory | up to 64k |
|
| Gemma 3 27B | Runs well | est. 13.5 tok/s10.8–16.3calibrated estimate ±20% | Q4_K_M | 17.8 / 44.8 GB | Unified memory | up to 128k |
|
| Qwen3.6 27B | Runs well | est. 13.4 tok/s10.8–16.1calibrated estimate ±20% | Q4_K_M | 17.9 / 44.8 GB | Unified memory | up to 32k |
|
| Muse Glimmer 30B | Runs well | est. 13.4 tok/s10.7–16.1calibrated estimate ±20% | Q4_K_M | 18.0 / 44.8 GB | Unified memory | up to 128k |
|
| Qwen3.5 27B | Runs well | est. 13.2 tok/s10.6–15.9calibrated estimate ±20% | Q4_K_M | 18.2 / 44.8 GB | Unified memory | up to 32k |
|
| Qwen3.8 27B | Runs well | est. 13.0 tok/s10.4–15.6calibrated estimate ±20% | Q4_K_M | 18.6 / 44.8 GB | Unified memory | up to 16k |
|
| Gemma 4 31B | Runs well | est. 12.3 tok/s9.8–14.7calibrated estimate ±20% | Q4_K_M | 19.6 / 44.8 GB | Unified memory | up to 8k |
|
| Kolibri 1 | Runs slowly | est. 55.6 tok/s39.0–72.3theoretical estimate ±30% | Q4_K_M | 49.9 GB RAM | CPU only | up to 256k |
|
| Granite 4.2 30B | Runs slowly | est. 11.7 tok/s9.4–14.1calibrated estimate ±20% | Q4_K_M | 20.4 / 44.8 GB | Unified memory | up to 64k |
|
| EXAONE 4.0 32B | Runs slowly | est. 11.7 tok/s9.4–14.1calibrated estimate ±20% | Q4_K_M | 20.5 / 44.8 GB | Unified memory | up to 128k |
|
| EXAONE 4.5 33B | Runs slowly | est. 11.3 tok/s9.1–13.6calibrated estimate ±20% | Q4_K_M | 21.2 / 44.8 GB | Unified memory | up to 256k |
|
| Qwen3 32B | Runs slowly | est. 10.7 tok/s8.5–12.8calibrated estimate ±20% | Q4_K_M | 22.5 / 44.8 GB | Unified memory | up to 32k |
|
| DeepSeek R1 Distill Qwen 32B | Runs slowly | est. 10.6 tok/s8.5–12.7calibrated estimate ±20% | Q4_K_M | 22.6 / 44.8 GB | Unified memory | up to 64k |
|
| Nanbeige4.3 33B | Runs slowly | est. 10.4 tok/s8.3–12.4calibrated estimate ±20% | Q4_K_M | 23.1 / 44.8 GB | Unified memory | up to 64k |
|
| Llama 3.3 70B | Runs slowly | est. 3.4 tok/s2.4–4.4theoretical estimate ±30% | Q4_K_M | 45.2 GB RAM | CPU only | up to 32k |
|
| Qwen3.5 122B-A10B | Won't runTry UD-Q2_K_XL (heavy quality loss): Runs great | est. 37.7 tok/s30.2–45.2calibrated estimate ±20% | UD-Q2_K_XL | 43.6 / 44.8 GB | Unified memory | up to 32k |
|
| Solar Open 100B | Won't runTry IQ4_XS: Runs slowly | est. 16.3 tok/s11.4–21.2theoretical estimate ±30% | IQ4_XS | 57.1 GB RAM | CPU only | up to 16k |
|
| DeepSeek V4 Flash 0731 | Won't run | — | MXFP4 | needs 155.6 GB | — | — |
|
| GLM-5.3 Flash | Won't run | — | Q4_K_M | needs 182.1 GB | — | — |
|
| gpt-oss-120b | Won't run | — | MXFP4 | needs 64.3 GB | — | — |
|
| Laguna S 2.1 | Won't run | — | Q4_K_M | needs 68.6 GB | — | — |
|
| Ling-3.0 Flash | Won't run | — | Q4_K_M | needs 77.7 GB | — | — |
|
| Solar Open 2 250B | Won't run | — | IQ4_XS | needs 137.2 GB | — | — |
|
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Verdict by memory size
| Model | 24 GB | 48 GB | 64 GB |
|---|---|---|---|
| MiniCPM5 1B | Runs great | Runs great | Runs great |
| Llama 3.2 1B | Runs great | Runs great | Runs great |
| Ling-3.0 Tiny | Runs great | Runs great | Runs great |
| EXAONE 4.0 1.2B | Runs great | Runs great | Runs great |
| LFM2.5 2.6B | Runs great | Runs great | Runs great |
| HyperCLOVA X SEED 1.5B | Runs great | Runs great | Runs great |
| MiniCPM5 2B | Runs great | Runs great | Runs great |
| Gemma 4 E2B | Runs great | Runs great | Runs great |
| Mellum2.1 12B-A2.5B Thinking | Runs great | Runs great | Runs great |
| MiniCPM-V 4.7 35B-A3B | Won't run | Runs great | Runs great |
| Spark-X2.5 4B | Runs great | Runs great | Runs great |
| Llama 3.2 3B | Runs great | Runs great | Runs great |
| gpt-oss-20b | Runs great | Runs great | Runs great |
| Nemotron 3.5 Lightning 30B-A3B | Won't run | Runs great | Runs great |
| Qwen3.5 4B | Runs great | Runs great | Runs great |
| Qwen3.6 35B-A3B | Won't run | Runs great | Runs great |
| Kanana 1.5 15.7B-A3B | Runs great | Runs great | Runs great |
| Qwen3.5 35B-A3B | Won't run | Runs great | Runs great |
| Gemma 4 E4B | Runs great | Runs great | Runs great |
| Qwen3 30B-A3B (2507) | Runs slowly | Runs great | Runs great |
| Gemma 4 26B-A4B | Runs slowly | Runs great | Runs great |
| Qwen3 8B | Runs great | Runs great | Runs great |
| K2-Horizon MoVA 36B-A4B | Won't run | Runs great | Runs great |
| DeepSeek R1 Distill Llama 8B | Runs great | Runs great | Runs great |
| Kanana 1.5 8B | Runs great | Runs great | Runs great |
| Llama 3.1 8B | Runs great | Runs great | Runs great |
| Nanbeige4.2 3B | Runs great | Runs great | Runs great |
| Qwen3.5 9B | Runs great | Runs great | Runs great |
| DeepSeek R1 0528 Qwen3 8B | Runs great | Runs great | Runs great |
| K2-Horizon 7B | Runs great | Runs great | Runs great |
| Gemma 4 12B | Runs great | Runs great | Runs great |
| Gemma 3 12B | Runs great | Runs great | Runs great |
| Phi-4 | Runs great | Runs great | Runs great |
| HyperCLOVA X SEED Think 14B | Runs well | Runs well | Runs well |
| Qwen3 14B | Runs well | Runs well | Runs well |
| DeepSeek R1 Distill Qwen 14B | Runs well | Runs well | Runs well |
| Mistral Small 3.2 24B | Runs well | Runs well | Runs well |
| Gemma 3 27B | Runs slowly | Runs well | Runs well |
| Qwen3.6 27B | Runs slowly | Runs well | Runs well |
| Muse Glimmer 30B | Runs slowly | Runs well | Runs well |
| Qwen3.5 27B | Runs slowly | Runs well | Runs well |
| Qwen3.8 27B | Runs slowly | Runs well | Runs well |
| Gemma 4 31B | Runs slowly | Runs well | Runs well |
| Kolibri 1 | Won't run | Won't run | Runs slowly |
| Granite 4.2 30B | Runs slowly | Runs slowly | Runs slowly |
| EXAONE 4.0 32B | Runs slowly | Runs slowly | Runs slowly |
| EXAONE 4.5 33B | Won't run | Runs slowly | Runs slowly |
| Qwen3 32B | Won't run | Runs slowly | Runs slowly |
| DeepSeek R1 Distill Qwen 32B | Won't run | Runs slowly | Runs slowly |
| Nanbeige4.3 33B | Won't run | Runs slowly | Runs slowly |
| Llama 3.3 70B | Won't run | Won't run | Runs slowly |
| Qwen3.5 122B-A10B | Won't run | Won't run | Won't run |
| Solar Open 100B | Won't run | Won't run | Won't run |
| DeepSeek V4 Flash 0731 | Won't run | Won't run | Won't run |
| GLM-5.3 Flash | Won't run | Won't run | Won't run |
| gpt-oss-120b | Won't run | Won't run | Won't run |
| Laguna S 2.1 | Won't run | Won't run | Won't run |
| Ling-3.0 Flash | Won't run | Won't run | Won't run |
| Solar Open 2 250B | Won't run | Won't run | Won't run |
Measured results on the Apple M5 Pro
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Model | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | llama.cpp | 512 | 1,340.0 | 67.6 | Metal, FA | github.com | 2026-09-24 |
| Qwen3 8B | Q4_K_M | llama.cpp | 512 | 1,223.0 | 45.4 | Metal | github.com | 2026-09-08 |
Frequently asked questions
What is the largest model that runs entirely on the Apple M5 Pro?
Kolibri 1 at IQ4_XS (a 43.63 GB file) fits entirely in 64 GB of unified memory with 8k context, at est. 57.5 tok/s (46.0–69.0, calibrated estimate ±20%).
How many local LLMs run well on the Apple M5 Pro?
At Q4_K_M with 8k context, with 64 GB of memory, out of 59 tracked models: 33 run great, 10 run well, 8 run slowly and 8 won't run.
Can the Apple M5 Pro run a 70B model like Llama 3.3 70B?
Runs slowly — CPU only, Q4_K_M: est. 3.4 tok/s (2.4–4.4, theoretical estimate ±30%).
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.