Apple · M4 Pro · Mini PC
Best AI Models for Mac Mini M4 Pro (24 GB)
Runs 14B models well and fits 35B MoE; dense 27–31B fits but is slow at 11 tok/s, and 70B doesn't fit in 20.5 GB.
| # | Model | Speed | llmrun Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 11.0 t/s | 11.0 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 11.0 t/s | 11.0 tok/s | Q4_K_M · 17.9 GB | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 49.1 t/s | 49.1 tok/s | Q4_K_M · 16.5 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 54.5 t/s | 54.5 tok/s | Q4_K_M · 13.1 GB | |
| 5 | LFM2.5 24B A2BLiquid AI Q4_K_M · 15.0 GB · 62.7 t/s | 62.7 tok/s | Q4_K_M · 15.0 GB |
| # | Model | Speed | Coding Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Q4_K_M · 16.5 GB · 49.1 t/s | 49.1 tok/s | Q4_K_M · 16.5 GB | |
| 2 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 54.5 t/s | 54.5 tok/s | Q4_K_M · 13.1 GB | |
| 3 | Q4_K_M · 7.9 GB · 36.1 t/s | 36.1 tok/s | Q4_K_M · 7.9 GB | |
| 4 | Qwen3.8 27BAlibaba Q4_K_M · 17.9 GB · 11.0 t/s | 11.0 tok/s | Q4_K_M · 17.9 GB | |
| 5 | LFM2.5 24B A2BLiquid AI Q4_K_M · 15.0 GB · 62.7 t/s | 62.7 tok/s | Q4_K_M · 15.0 GB |
| # | Model | Speed | Reasoning Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Q4_K_M · 16.5 GB · 49.1 t/s | 49.1 tok/s | Q4_K_M · 16.5 GB | |
| 2 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 30.0 t/s | 30.0 tok/s | Q4_K_M · 6.6 GB | |
| 3 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 54.5 t/s | 54.5 tok/s | Q4_K_M · 13.1 GB | |
| 4 | Qwen3 8BAlibaba Q4_K_M · 7.6 GB · 34.6 t/s | 34.6 tok/s | Q4_K_M · 7.6 GB | |
| 5 | LFM2.5 24B A2BLiquid AI Q4_K_M · 15.0 GB · 62.7 t/s | 62.7 tok/s | Q4_K_M · 15.0 GB |
| # | Model | Speed | Long context Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 30.0 t/s | 30.0 tok/s | Q4_K_M · 6.6 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 58.6 t/s | 58.6 tok/s | Q4_K_M · 3.4 GB | |
| 3 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 54.5 t/s | 54.5 tok/s | Q4_K_M · 13.1 GB | |
| 4 | Mistral Nemo Instruct 2407Mistral AI Q4_K_M · 11.0 GB · 23.7 t/s | 23.7 tok/s | Q4_K_M · 11.0 GB | |
| 5 | LFM2.5 24B A2BLiquid AI Q4_K_M · 15.0 GB · 62.7 t/s | 62.7 tok/s | Q4_K_M · 15.0 GB |
| # | Model | Speed ↓ | llmrun Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 58.6 t/s | 58.6 tok/s | Q4_K_M · 3.4 GB | |
| 2 | LFM2.5 24B A2BLiquid AI Q4_K_M · 15.0 GB · 62.7 t/s | 62.7 tok/s | Q4_K_M · 15.0 GB | |
| 3 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 54.5 t/s | 54.5 tok/s | Q4_K_M · 13.1 GB | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 49.1 t/s | 49.1 tok/s | Q4_K_M · 16.5 GB | |
| 5 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 30.0 t/s | 30.0 tok/s | Q4_K_M · 6.6 GB |
| # | Model | Speed | Vision Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Q4_K_M · 16.5 GB · 49.1 t/s | 49.1 tok/s | Q4_K_M · 16.5 GB | |
| 2 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 30.0 t/s | 30.0 tok/s | Q4_K_M · 6.6 GB | |
| 3 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 58.6 t/s | 58.6 tok/s | Q4_K_M · 3.4 GB | |
| 4 | Gemma 3 12B ITGoogle Q4_K_M · 10.9 GB · 23.8 t/s | 23.8 tok/s | Q4_K_M · 10.9 GB | |
| 5 | Q4_K_M · 6.5 GB · 30.8 t/s | 30.8 tok/s | Q4_K_M · 6.5 GB |
The Mac Mini M4 Pro is one of the fastest 24 GB machines: it runs 14B models well and fits 35B MoE. Among 24 GB machines, only the M5 Pro write faster. macOS keeps about 3.5 GB, leaving 20.5 GB for models. With 17 TFLOPS it reads long prompts faster than most 24 GB machines: a 4K-token prompt takes 66 s. At its $1,599 MSRP it gives 18.8 tok/s per $1,000, 2nd of the 8 machines compared here. M4 Pro generation, released 2024; the newer M5 Pro writes 13% faster with 1.5× the compute.
Mac Mini M4 Pro strengths and weaknesses
Strengths
- Runs 35B MoE models: Qwen3.6 35B A3B at 46.3 tok/s in 19.8 GB
- Low power draw: 140 W at full load for the whole machine
Weaknesses
- 70B models don't fit: Llama 3.3 70B Instruct needs 51.3 GB at Q4 vs 20.5 GB usable here
- Slow on large dense models: Qwen3.8 27B at 11.0 tok/s
- Slow prompt reading: A 32K-token prompt on Qwen3.8 27B waits 8.8 min for the first token
- No CUDA: vLLM, AWQ, GPTQ and FP8 builds don't run; use GGUF or MLX
Memory for models
20.5 GB of 24 GB70B models need 51.3 GB at Q4, more than the 20.5 GB available. The 48 GB config fits Qwen3.8 27B →Or rent an A100 (80 GB) from $1.11/hr →
Speed per dollar and per watt
On Qwen3.5 9B (9B params) at Q4_K_M. New means a shop price, either a street quote or the maker's own price; used is compared with used. The last two columns are the speed you get per $1,000 of price and per 100 W of power; higher is better. The Mac Mini M4 Pro has no current new price, so it is shown at its MSRP.
| Machine | New price | tok/s | tok/s per $1,000 | tok/s per 100 W |
|---|---|---|---|---|
RTX 509032 GB | about $5,376 | 183 | 34.1 | 31.8 |
Mac Mini M4 Pro24 GB | about $1,599 | 30.0 | 18.8 | 21.4 |
Mac Studio M4 Max36 GB | about $2,499 | 45.1 | 18.0 | 31.1 |
| about $1,999 | 30.0 | 15.0 | – | |
| about $2,299 | 33.8 | 14.7 | – | |
Mac Mini M4 Pro48 GB | about $2,199 | 30.0 | 13.6 | 21.4 |
| about $2,499 | 30.0 | 12.0 | – | |
MacBook Air 13" M424 GB | about $1,399 | 13.2 | 9.4 | – |
| Machine | Used price | tok/s | tok/s per $1,000 | tok/s per 100 W |
|---|---|---|---|---|
RTX 409024 GBBest used value | about $3,000 | 103 | 34.3 | 22.9 |
RTX 509032 GB | about $6,400 | 183 | 28.6 | 31.8 |
Mac Studio M4 Max36 GB | about $2,000 | 45.1 | 22.6 | 31.1 |
| about $2,109 | 33.8 | 16.0 | – |
Cost per token: the Mac Mini M4 Pro against 24–64 GB machines and GPUs
On Qwen3.5 9B (9B params, Q4_K_M, 4K context). Speed, cost per token and power per token depend on the model.
- Machine, MSRP
- $3.38
- Power
- $0.19
Renting costs as much as buying after ≈3,500 hours (≈2.4 years at 4 h a day).
Each dot is a machine or GPU; hover or tap it for its numbers, click to open it. Machines without a price for the chosen basis are left out. Machines without a published power figure are left out of Per watt and costed without power. How we calculate cost per token
Other memory sizes of the same machine first, then machines and GPUs people weigh against it, each with the one reason it is worth a look. Speeds on Qwen3.5 9B.
Output speed depends on memory bandwidth (273 GB/s). Grey rows need more than the 20.5 GB available.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB30.0tok/s
- 14BQwen3 14B · Q4_K_M · 11.9 GB20.1tok/s
- 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB11.0tok/s
- 46.3tok/s
- 70BLlama 3.3 70B Instruct · not at 4-bitneeds 51.3 GB
- 120B MoEGPT OSS 120B · not at 4-bitneeds 70.8 GB
Estimated speeds: Q4_K_M, 4K context, one user. How we estimate
Pro generations: writing vs compute
M1 Pro → M5 ProWriting speed follows memory bandwidth and barely moved: 22.0 to 33.8 tok/s since the M1 Pro. Compute, which sets how fast prompts are read, grew 2.8×. Indexed to the M1 Pro; speeds on Qwen3.5 9B, same 16 GB.
All 1125 models on the Mac Mini M4 Pro
Featured onlyBest quant that fits in 20.5 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works
| Model | Quant | VRAM | Speed | Context | Score | Fit |
|---|---|---|---|---|---|---|
Typhoon Ocr 3BVision | Q4_K_M | 3.2 GB | 72.7 t/s | 128K | S86, Runs great | |
OLMo 1B HFChat | Q4_K_M | 1.3 GB | 150.5 t/s | 2K | S97, Runs great | |
VLM2Vec FullChat | BF16 | 15.0 GB | 20.3 t/s | 131K | B60, Decent | |
Qwen3 30B A3B.w8a8ChatFunctions | Q4_K_M | 19.4 GB | 49.0 t/s | 41K | B56, Decent | |
| Q4_K_M | 0.1 GB | 1470.0 t/s | 2K | S100, Runs great | ||
GPT Neo 125MChat | BF16 | 0.3 GB | 579.1 t/s | 2K | S100, Runs great | |
DeepSeek R1 Distill Qwen 14B Abliterated v2ChatReasoning | Q4_K_M | 12.4 GB | 20.0 t/s | 131K | B60, Decent | |
LFM2 VL 1.6BVision | Q4_K_M | 1.5 GB | 144.8 t/s | 128K | S96, Runs great | |
Granite Switch 4.1 30BChatFunctions | Q3_K_L | 20.1 GB | 11.8 t/s | 131K | D23, Barely runs | |
Bloom 560MChat | Q4_K_M | 1.6 GB | 516.5 t/s | — | S100, Runs great | |
| Q4_K_M | 0.6 GB | 444.4 t/s | 8K | S100, Runs great | ||
PLLuM 12B ChatChat | Q4_K_M | 11.0 GB | 23.7 t/s | 131K | B64, Decent | |
| BF16 | 8.6 GB | 29.0 t/s | — | A69, Runs well | ||
| Q4_K_M | 11.0 GB | 23.7 t/s | 1024K | B64, Decent | ||
Huihui Qwen3 Coder 30B A3B Instruct AbliteratedChatCodeFunctions | Q4_K_M | 19.4 GB | 49.1 t/s | 262K | B56, Decent |
Specs
- Chip
- M4 Pro
- CPU cores
- 14
- GPU cores
- 20
- Unified memory
- 24 GB
- For models
- 20.5 GB (macOS keeps 3.5 GB)
- Memory bandwidth
- 273 GB/s
- FP16 (dense)
- 17.2 TFLOPS
- Neural Engine
- 38 TOPS
- Power
- 140 W max
- Released
- 2024-11-08
- MSRP
- $1,599
Mac Mini M4 Pro FAQ
Is the Mac Mini M4 Pro (24 GB) worth it for local LLMs in 2026?
At its $1,599 MSRP it gives 18.8 tok/s per $1,000 on Qwen3.5 9B, 2nd of the 8 machines compared here. It runs 14B models well and fits 35B MoE; dense 27–31B fits but is slow at 11 tok/s, and 70B doesn't fit in 20.5 GB.
How much memory is available for AI on the Mac Mini M4 Pro (24 GB)?
About 20.5 GB of the 24 GB. macOS keeps roughly 3.5 GB for itself and other apps.
What is the largest model the Mac Mini M4 Pro (24 GB) can run?
Qwen3.6 35B A3B (36.0B) is the largest at 4-bit or better: at IQ4_XS it uses 19.8 GB, 97% of the available memory, and runs at about 46.3 tok/s.
Can the Mac Mini M4 Pro (24 GB) run Qwen3.8 27B?
Yes. Qwen3.8 27B fits at Q4_K_M, using 17.9 GB of the 20.5 GB available, and it is the best model that runs well on this machine.
Can the Mac Mini M4 Pro (24 GB) run 70B models?
No. Llama 3.3 70B Instruct needs 51.3 GB at Q4, more than the 20.5 GB available.
How long does the Mac Mini M4 Pro (24 GB) take to read a long prompt?
On Qwen3.8 27B, about 66 s for a 4K-token prompt and 8.8 min for a 32K-token document before the first word appears.
Is renting a GPU cheaper than buying the Mac Mini M4 Pro?
An RTX 5090 (32 GB) rents from $0.46/hr. Renting costs as much as buying after ≈3,500 hours (≈2.4 years at 4 h a day).