What LLM Can a Mac mini Run? 16 GB to 64 GB
Short answer. What a Mac mini can run depends on its memory: macOS gives the GPU about three quarters of it by default. A 16 GB Mac mini runs models up to about 8–12B parameters at 4-bit (Gemma 4 12B, Qwen 3.5 9B, Llama 3.1 8B); 24 GB adds gpt-oss 20B and the 14B reasoning models; 32 GB is the big step, opening the popular 24–35B models such as Qwen3-Coder 30B and Qwen 3.8 27B; 48 GB adds dense 30–32B models; 64 GB runs 70B models only tightly. Speed depends on the chip: the M4 Pro and M5 Pro have more than twice the memory bandwidth of the M4. Choose your Mac mini below to see every model that runs on it and what renting one costs.
Key facts
- Mac mini memory options: M4 and M6 with 16, 24 or 32 GB; M4 Pro and M5 Pro with 24, 48 or 64 GB (Apple, Apple).
- Memory bandwidth per Apple: M4 120 GB/s, M6 153 GB/s (16 GB) or 170 GB/s (24/32 GB), M4 Pro 273 GB/s, M5 Pro 307 GB/s.
- The cheapest way to run 30B mixture-of-experts models such as Qwen3-Coder is a 32 GB M4 or M6: from €129 a month (MacDuty, M4 32 GB).
- gpt-oss 20B is the largest popular model that fits 24 GB: about 16 GB.
- No Mac mini runs gpt-oss 120B or Qwen 3.5 122B; 70B models fit only the 64 GB Mac mini, tightly.
- For image and video generation the chip matters more than memory bandwidth: a 5-second MiniMax H3 clip took 17 min 43 s on a Mac mini M4 Pro and 5 min 29 s on an M5 Pro (vpipe developer); Draw Things measured the M6 1.06–1.58× faster than an M4 Pro Mac mini.
- Figures are for 4-bit (Q4) builds with a 32K context, from our catalog of 27 local language models checked 2 November 2026; speeds are estimates, not measurements.
Which LLMs run on your Mac mini?
Choose a Mac mini configuration from our price tables. The list shows every model size that runs on it, largest first, with the memory it needs and the estimated speed on that chip; filter by task, change quantization or context, and open "Details" for the model's install command and all cloud Mac offers that can run it.
M4, 16 GB: about 12 GB for models by default, 120 GB/s memory bandwidth (Apple). 44 model sizes run on it, largest first. 56 offers from €74/mo.
| Model | Memory | Fit | Speed (est.) | Details |
|---|---|---|---|---|
| FLUX.2 [klein] 3.9BImage generation | from 16 GB | Fits | see reports | |
| Z-Image-Turbo 6.2BImage generation | from 16 GB | Fits | see reports | |
| Krea 2 Turbo 12.8BImage generation | from 16 GB | Fits | see reports | |
| Stable Diffusion XL 2.6BImage generation | from 16 GB | Fits | see reports | |
| MiniMax H3 33.1BVideo generation | from 16 GB | Fits | see reports | |
| LTX-2.5 22BVideo generation | from 16 GB | Fits | see reports | |
| Whisper large-v3-turbo 809MSpeech to text | from 16 GB | Fits | see reports | |
| Parakeet TDT 0.6B v3 627MSpeech to text | from 16 GB | Fits | see reports | |
| Qwen3-ASR 2.3BSpeech to text | from 16 GB | Fits | see reports | |
| Kokoro-82MText to speech | from 16 GB | Fits | see reports | |
| Qwen3-TTS 1.9BText to speech | from 16 GB | Fits | see reports | |
| Chatterbox 0.5BText to speech | from 16 GB | Fits | see reports | |
| Phonon 41MText to speech | from 16 GB | Fits | see reports | |
| Qwen3-VL 8BImages · Documents and OCR · AI agents | 12 GB | Fits | ≈ 10–15 tok/s | |
| Granite 4.2 8BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 12 GB | Fits | ≈ 11–17 tok/s | |
| TranslateGemma 12BTranslation · Many languages | 12 GB | Fits | ≈ 7–11 tok/s | |
| Gemma 4 12BChat and writing · Reasoning and math · AI agents · Coding · Images | 11 GB | Fits | ≈ 8–11 tok/s | |
| Bonsai 2 27B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 11 GB | Fits | ≈ 7–11 tok/s | |
| Gemma 3n E4BChat and writing · Images · Many languages | 11 GB | Fits | ≈ 6–9 tok/s | |
| Gemma 4 E4BChat and writing · Reasoning and math · AI agents · Coding · Images | 11 GB | Fits | ≈ 6–9 tok/s | |
| DeepSeek-R1 (distilled) 8B (0528)Reasoning and math · Coding | 11 GB | Fits | ≈ 12–17 tok/s | |
| Qwen3 Embedding 8BSearch your documents (RAG) | 10 GB | Fits | ≈ 13–19 tok/s | |
| Llama 3.1 8BChat and writing · Many languages · Documents and OCR · Coding | 10 GB | Fits | ≈ 12–18 tok/s | |
| Qwen 3.5 9BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 9.6 GB | Fits | ≈ 8–12 tok/s | |
| Nemotron 3.5 Bolt E4BChat and writing · Coding · AI agents · Images | 9.4 GB | Fits | ≈ 8–11 tok/s | |
| Gemma 3n E2BChat and writing · Images · Many languages | 9.2 GB | Fits | ≈ 8–12 tok/s | |
| Qwen3-VL 4BImages · Documents and OCR · AI agents | 9.0 GB | Fits | ≈ 18–27 tok/s | |
| Gemma 4 E2BChat and writing · Reasoning and math · AI agents · Coding · Images | 8.7 GB | Fits | ≈ 8–12 tok/s | |
| Ornith 1.5 9BCoding · AI agents | 8.6 GB | Fits | ≈ 9–14 tok/s | |
| Qwen3 Embedding 4BSearch your documents (RAG) | 8.2 GB | Fits | ≈ 24–36 tok/s | |
| DeepSeek-R1 (distilled) 7BReasoning and math · Coding | 7.5 GB | Fits | ≈ 13–19 tok/s | |
| Llama 3.2 3BChat and writing · Documents and OCR · Many languages | 6.7 GB | Fits | ≈ 30–45 tok/s | |
| Qwen3-VL 2BImages · Documents and OCR · AI agents | 6.6 GB | Fits | ≈ 32–47 tok/s | |
| Qwen 3.5 4BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 6.0 GB | Fits | ≈ 15–23 tok/s | |
| Granite 4.2 3BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 5.8 GB | Fits | ≈ 27–41 tok/s | |
| GLM-OCR 1.3BDocuments and OCR · Images | 5.8 GB | Fits | ≈ 22–33 tok/s | |
| Qwen3 Embedding 0.6BSearch your documents (RAG) | 5.3 GB | Fits | ≈ 94–141 tok/s | |
| Bonsai 2 9B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 5.2 GB | Fits | ≈ 19–28 tok/s | |
| TranslateGemma 4BTranslation · Many languages | 5.1 GB | Fits | ≈ 18–27 tok/s | |
| Qwen 3.5 2BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 4.5 GB | Fits | ≈ 19–29 tok/s | |
| Llama 3.2 1BChat and writing · Documents and OCR · Many languages | 3.3 GB | Fits | ≈ 46–69 tok/s | |
| LFM3 1.7BAI agents · Chat and writing | 3.1 GB | Fits | ≈ 35–53 tok/s | |
| DeepSeek-R1 (distilled) 1.5BReasoning and math · Coding | 3.0 GB | Fits | ≈ 55–82 tok/s | |
| Qwen 3.5 0.8BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 2.7 GB | Fits | ≈ 46–69 tok/s |
Gemma 4 E2B by Google DeepMind
Google's open multimodal family with a configurable thinking mode. The developer positions it for reasoning, agentic workflows, coding and image understanding; the small E2B/E4B sizes are built for laptops.
- Tasks: Everyday assistant · Step-by-step reasoning, math and logic · AI agents that call tools and finish multi-step tasks · Programming help and coding agents · Understands images and screenshots
- Input: text, image · Max context: 128K · Licence: Apache 2.0
- Download: Ollama · Hugging Face (original)
8.7 GB memory needed
model file 7.5 GB + context 0.2 GB (32K tokens) + runtime ≈ 1 GB
ollama run gemma4:e2bWhich cloud Mac mini runs it?
M1 · 8 GBDoes not fit–18 offers, from €29/mo
M2 · 8 GBDoes not fit–15 offers, from €54/mo
M1 · 16 GBFits–15 offers, from €75/mo
M2 · 16 GBFits≈ 7–10 tok/s15 offers, from €89/mo
M2 Pro · 16 GBFits≈ 13–20 tok/s10 offers, from €128/mo
M4 · 16 GBFits≈ 8–12 tok/s56 offers, from €74/mo
M6 · 16 GBFits≈ 10–15 tok/s26 offers, from €101/mo
M2 · 24 GBFits≈ 7–10 tok/s7 offers, from €86/mo
M4 · 24 GBFits≈ 8–12 tok/s13 offers, from €117/mo
M4 Pro · 24 GBFits≈ 18–27 tok/s10 offers, from €118/mo
M5 Pro · 24 GBFits≈ 20–31 tok/s9 offers, from €178/mo
M6 · 24 GBFits≈ 11–17 tok/s10 offers, from €139/mo
M2 Pro · 32 GBFits≈ 13–20 tok/s6 offers, from €180/mo
M4 · 32 GBFits≈ 8–12 tok/s10 offers, from €129/mo
M6 · 32 GBFits≈ 11–17 tok/s15 offers, from €135/mo
M4 Pro · 48 GBFits≈ 18–27 tok/s9 offers, from €271/mo
M5 Pro · 48 GBFits≈ 20–31 tok/s8 offers, from €225/mo
M4 Pro · 64 GBFits≈ 18–27 tok/s11 offers, from €239/mo
M5 Pro · 64 GBFits≈ 20–31 tok/s13 offers, from €226/mo
243 cloud Mac offers that can run it, cheapest first
| Provider | Mac | Where | Speed (est.) | Per month |
|---|---|---|---|---|
| GreenMini | M4, 16 GB / 256 GB | NetherlandsEurope | ≈ 8–12 tok/s | from €74 |
| MacDuty | M1, 16 GB / 256 GB | RomaniaEurope | – | from €75 |
| HostMyApple | M1, 16 GB / 256 GB | Atlanta, GA, US, Denver, CO, US, Amsterdam, NLEurope | – | €77 |
| Mac mini VaultM1 256GB | M1, 16 GB / 256 GBout of stock | Milwaukee, WI, US, Phoenix, AZ, USNorth America | – | €77 |
| HostMyApple | M1, 16 GB / 256 GB | Atlanta, GA, US, Denver, CO, US, Amsterdam, NLNorth America | – | €77 |
| My Remote Mac | M4, 16 GB / 256 GB | Nantes, FREurope | ≈ 8–12 tok/s | €77 |
| xaweho | M4, 16 GB / 256 GBout of stock | GermanyEurope | ≈ 8–12 tok/s | €79 |
| OakHostMac M1.M | M1, 16 GB / 256 GBout of stock | Nuremberg, DE, Bamberg, DEEurope | – | €80 |
| MacHiveMac mini M4 | M4, 16 GB / 512 GB | United KingdomEurope | ≈ 8–12 tok/s | €83 |
| OakHostMac M1.L | M1, 16 GB / 1 TB | Nuremberg, DE, Bamberg, DEEurope | – | €85 |
| Mac mini VaultM2 256GB | M2, 24 GB / 256 GBout of stock | Milwaukee, WI, US, Phoenix, AZ, USNorth America | ≈ 7–10 tok/s | €86 |
| MacDuty | M2, 16 GB / 256 GB | RomaniaEurope | ≈ 7–10 tok/s | from €89 |
| ZoneMacMac mini M4 | M4, 16 GB / 256 GB | Singapore, Tokyo, JP, Seoul, KR, Hong KongAsia and Australia | ≈ 8–12 tok/s | €89 |
| RoundfleetM4-S | M4, 16 GB / 256 GB | Toulouse, FREurope | ≈ 8–12 tok/s | €90 |
| My Remote Mac | M4, 16 GB / 512 GB | Nantes, FREurope | ≈ 8–12 tok/s | €90 |
The list also includes image, video, speech, music and 3D models: for them it shows the smallest Mac memory reported to work, and speeds measured by developers and users instead of our estimate.
The price tables link here: "which LLMs?" next to each plan's memory opens this tool with that Mac selected. How memory and speed are calculated is explained in Which Mac mini for a local LLM?; models grouped by task are in Best local LLM by task.
What LLM can a Mac mini with 16 GB run?
About 12 GB is available to the GPU, enough for the small and mid-size models: general assistants up to 12B, the 8B reasoning distills, Ornith 1.5 9B for coding, Qwen3-VL 8B for images, GLM-OCR for documents, Bonsai 2 27B (ternary weights, so a 27B model fits; speed is our estimate), TranslateGemma 12B and the embedding models for search. A 16 GB Mac mini comes with the M4 or the M6; the M6 is faster thanks to its higher bandwidth.
| Model | Memory needed | Speed on M4 | Speed on M6 |
|---|---|---|---|
| FLUX.2 [klein] 3.9BImage generation | from 16 GB Macreported | see reports | see reports |
| Z-Image-Turbo 6.2BImage generation | from 16 GB Macreported | see reports | see reports |
| Krea 2 Turbo 12.8BImage generation | from 16 GB Macreported | see reports | see reports |
| Stable Diffusion XL 2.6BImage generation | from 16 GB Macreported | see reports | see reports |
| MiniMax H3 33.1BVideo generation | from 16 GB Macreported | see reports | see reports |
| LTX-2.5 22BVideo generation | from 16 GB Macreported | see reports | see reports |
| Whisper large-v3-turbo 809MSpeech to text | from 16 GB Macreported | see reports | see reports |
| Parakeet TDT 0.6B v3 627MSpeech to text | from 16 GB Macreported | see reports | see reports |
| Qwen3-ASR 2.3BSpeech to text | from 16 GB Macreported | see reports | see reports |
| Kokoro-82MText to speech | from 16 GB Macreported | see reports | see reports |
| Qwen3-TTS 1.9BText to speech | from 16 GB Macreported | see reports | see reports |
| Chatterbox 0.5BText to speech | from 16 GB Macreported | see reports | see reports |
| Phonon 41MText to speech | from 16 GB Macreported | see reports | see reports |
| Qwen3-VL 8BImages · Documents and OCR · AI agents | 12 GB | ≈ 10–15 tok/s | ≈ 12–19 tok/s |
| Granite 4.2 8BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 12 GB | ≈ 11–17 tok/s | ≈ 14–22 tok/s |
| TranslateGemma 12BTranslation · Many languages | 12 GB | ≈ 7–11 tok/s | ≈ 9–14 tok/s |
| Gemma 4 12BChat and writing · Reasoning and math · AI agents · Coding · Images | 11 GB | ≈ 8–11 tok/s | ≈ 10–14 tok/s |
| Bonsai 2 27B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 11 GB | ≈ 7–11 tok/s | ≈ 9–14 tok/s |
| Gemma 3n E4BChat and writing · Images · Many languages | 11 GB | ≈ 6–9 tok/s | ≈ 8–12 tok/s |
| DeepSeek-R1 (distilled) 8B (0528)Reasoning and math · Coding | 11 GB | ≈ 12–17 tok/s | ≈ 15–22 tok/s |
| Qwen3 Embedding 8BSearch your documents (RAG) | 10 GB | ≈ 13–19 tok/s | ≈ 16–24 tok/s |
| Llama 3.1 8BChat and writing · Many languages · Documents and OCR · Coding | 10 GB | ≈ 12–18 tok/s | ≈ 16–23 tok/s |
| Nemotron 3.5 Bolt E4BChat and writing · Coding · AI agents · Images | 9.4 GB | ≈ 8–11 tok/s | ≈ 10–15 tok/s |
| Gemma 3n E2BChat and writing · Images · Many languages | 9.2 GB | ≈ 8–12 tok/s | ≈ 10–15 tok/s |
| Qwen3-VL 4BImages · Documents and OCR · AI agents | 9.0 GB | ≈ 18–27 tok/s | ≈ 23–35 tok/s |
| Qwen 3.5 9BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 8.6 GB | ≈ 9–14 tok/s | ≈ 12–17 tok/s |
| Ornith 1.5 9BCoding · AI agents | 8.6 GB | ≈ 9–14 tok/s | ≈ 12–17 tok/s |
| Qwen3 Embedding 4BSearch your documents (RAG) | 8.2 GB | ≈ 24–36 tok/s | ≈ 31–46 tok/s |
| Gemma 4 E4BChat and writing · Reasoning and math · AI agents · Coding · Images | 8.1 GB | ≈ 9–14 tok/s | ≈ 12–17 tok/s |
| DeepSeek-R1 (distilled) 7BReasoning and math · Coding | 7.5 GB | ≈ 13–19 tok/s | ≈ 16–24 tok/s |
| Llama 3.2 3BChat and writing · Documents and OCR · Many languages | 6.7 GB | ≈ 30–45 tok/s | ≈ 38–57 tok/s |
| Qwen3-VL 2BImages · Documents and OCR · AI agents | 6.6 GB | ≈ 32–47 tok/s | ≈ 40–60 tok/s |
| Gemma 4 E2BChat and writing · Reasoning and math · AI agents · Coding · Images | 5.8 GB | ≈ 13–20 tok/s | ≈ 17–25 tok/s |
| Granite 4.2 3BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 5.8 GB | ≈ 27–41 tok/s | ≈ 35–52 tok/s |
| GLM-OCR 1.3BDocuments and OCR · Images | 5.8 GB | ≈ 22–33 tok/s | ≈ 28–43 tok/s |
| Qwen 3.5 4BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 5.3 GB | ≈ 18–27 tok/s | ≈ 23–35 tok/s |
| Qwen3 Embedding 0.6BSearch your documents (RAG) | 5.3 GB | ≈ 94–141 tok/s | ≈ 120–179 tok/s |
| Bonsai 2 9B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 5.2 GB | ≈ 19–28 tok/s | ≈ 24–36 tok/s |
| TranslateGemma 4BTranslation · Many languages | 5.1 GB | ≈ 18–27 tok/s | ≈ 23–35 tok/s |
| Qwen 3.5 2BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 3.3 GB | ≈ 32–47 tok/s | ≈ 40–60 tok/s |
| DeepSeek-R1 (distilled) 1.5BReasoning and math · Coding | 3.0 GB | ≈ 55–82 tok/s | ≈ 70–104 tok/s |
| Llama 3.2 1BChat and writing · Documents and OCR · Many languages | 2.9 GB | ≈ 74–111 tok/s | ≈ 94–142 tok/s |
| Qwen 3.5 0.8BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 2.7 GB | ≈ 46–69 tok/s | ≈ 59–88 tok/s |
| LFM3 1.7BAI agents · Chat and writing | 2.5 GB | ≈ 55–82 tok/s | ≈ 70–104 tok/s |
Is 24 GB enough for local LLMs?
24 GB gives the GPU about 18 GB, which adds only a few models: gpt-oss 20B, which runs quickly because only 3.6B of its parameters are active, plus the DeepSeek-R1 14B distill and Phi-4. The popular 24–35B models need about 20–24 GB with a 32K context, so they do not fit. 24 GB is available on all four current chips; on an M4 Pro or M5 Pro it is the entry configuration.
| Model | Memory needed | Speed on M4 | Speed on M6 | Speed on M4 Pro | Speed on M5 Pro |
|---|---|---|---|---|---|
| FLUX.1 [dev] 11.9BImage generation | from 24 GB Macreported | see reports | see reports | see reports | see reports |
| Qwen-Image 2.1 7.1BImage generation | from 24 GB Macreported | see reports | see reports | see reports | see reports |
| ACE-Step 1.5 2B+Music | from 24 GB Macreported | see reports | see reports | see reports | see reports |
| TRELLIS.2 4B3D models | from 24 GB Macreported | see reports | see reports | see reports | see reports |
| Hunyuan3D 2.1 3.3B3D models | from 24 GB Macreported | see reports | see reports | see reports | see reports |
| DeepSeek-R1 (distilled) 14BReasoning and math · Coding | 16 GB | ≈ 7–10 tok/s | ≈ 9–14 tok/s | ≈ 15–23 tok/s | ≈ 17–26 tok/s |
| FLUX.2 [klein] 3.9BImage generation | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Z-Image-Turbo 6.2BImage generation | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Krea 2 Turbo 12.8BImage generation | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Stable Diffusion XL 2.6BImage generation | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| MiniMax H3 33.1BVideo generation | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| LTX-2.5 22BVideo generation | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Whisper large-v3-turbo 809MSpeech to text | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Parakeet TDT 0.6B v3 627MSpeech to text | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Qwen3-ASR 2.3BSpeech to text | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Kokoro-82MText to speech | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Qwen3-TTS 1.9BText to speech | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Chatterbox 0.5BText to speech | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| Phonon 41MText to speech | from 16 GB Macreported | see reports | see reports | see reports | see reports |
| gpt-oss 20B (MoE, 3.6B active)Reasoning and math · AI agents · Coding | 16 GB | ≈ 25–37 tok/s | ≈ 35–53 tok/s | ≈ 57–85 tok/s | ≈ 64–95 tok/s |
| Phi-4 14BReasoning and math · Chat and writing | 13 GB | ≈ 7–10 tok/s | ≈ 9–14 tok/s | ≈ 15–23 tok/s | ≈ 17–25 tok/s |
| Qwen3-VL 8BImages · Documents and OCR · AI agents | 12 GB | ≈ 10–15 tok/s | ≈ 14–21 tok/s | ≈ 22–33 tok/s | ≈ 25–37 tok/s |
| Granite 4.2 8BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 12 GB | ≈ 11–17 tok/s | ≈ 16–24 tok/s | ≈ 26–39 tok/s | ≈ 29–43 tok/s |
| TranslateGemma 12BTranslation · Many languages | 12 GB | ≈ 7–11 tok/s | ≈ 10–16 tok/s | ≈ 17–25 tok/s | ≈ 19–28 tok/s |
| Gemma 4 12BChat and writing · Reasoning and math · AI agents · Coding · Images | 11 GB | ≈ 8–11 tok/s | ≈ 11–16 tok/s | ≈ 17–26 tok/s | ≈ 19–29 tok/s |
| Bonsai 2 27B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 11 GB | ≈ 7–11 tok/s | ≈ 10–15 tok/s | ≈ 16–25 tok/s | ≈ 18–28 tok/s |
| Gemma 3n E4BChat and writing · Images · Many languages | 11 GB | ≈ 6–9 tok/s | ≈ 9–13 tok/s | ≈ 14–21 tok/s | ≈ 16–23 tok/s |
| DeepSeek-R1 (distilled) 8B (0528)Reasoning and math · Coding | 11 GB | ≈ 12–17 tok/s | ≈ 16–25 tok/s | ≈ 26–39 tok/s | ≈ 30–44 tok/s |
| Qwen3 Embedding 8BSearch your documents (RAG) | 10 GB | ≈ 13–19 tok/s | ≈ 18–27 tok/s | ≈ 29–44 tok/s | ≈ 33–49 tok/s |
| Llama 3.1 8BChat and writing · Many languages · Documents and OCR · Coding | 10 GB | ≈ 12–18 tok/s | ≈ 17–26 tok/s | ≈ 28–42 tok/s | ≈ 31–47 tok/s |
| Nemotron 3.5 Bolt E4BChat and writing · Coding · AI agents · Images | 9.4 GB | ≈ 8–11 tok/s | ≈ 11–16 tok/s | ≈ 17–26 tok/s | ≈ 19–29 tok/s |
| Gemma 3n E2BChat and writing · Images · Many languages | 9.2 GB | ≈ 8–12 tok/s | ≈ 11–16 tok/s | ≈ 18–26 tok/s | ≈ 20–30 tok/s |
| Qwen3-VL 4BImages · Documents and OCR · AI agents | 9.0 GB | ≈ 18–27 tok/s | ≈ 26–39 tok/s | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen 3.5 9BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 8.6 GB | ≈ 9–14 tok/s | ≈ 13–19 tok/s | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| Ornith 1.5 9BCoding · AI agents | 8.6 GB | ≈ 9–14 tok/s | ≈ 13–19 tok/s | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| Qwen3 Embedding 4BSearch your documents (RAG) | 8.2 GB | ≈ 24–36 tok/s | ≈ 34–51 tok/s | ≈ 55–82 tok/s | ≈ 61–92 tok/s |
| Gemma 4 E4BChat and writing · Reasoning and math · AI agents · Coding · Images | 8.1 GB | ≈ 9–14 tok/s | ≈ 13–19 tok/s | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| DeepSeek-R1 (distilled) 7BReasoning and math · Coding | 7.5 GB | ≈ 13–19 tok/s | ≈ 18–27 tok/s | ≈ 29–44 tok/s | ≈ 33–49 tok/s |
| Llama 3.2 3BChat and writing · Documents and OCR · Many languages | 6.7 GB | ≈ 30–45 tok/s | ≈ 43–64 tok/s | ≈ 68–102 tok/s | ≈ 77–115 tok/s |
| Qwen3-VL 2BImages · Documents and OCR · AI agents | 6.6 GB | ≈ 32–47 tok/s | ≈ 45–67 tok/s | ≈ 72–108 tok/s | ≈ 81–121 tok/s |
| Gemma 4 E2BChat and writing · Reasoning and math · AI agents · Coding · Images | 5.8 GB | ≈ 13–20 tok/s | ≈ 18–28 tok/s | ≈ 30–45 tok/s | ≈ 33–50 tok/s |
| Granite 4.2 3BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 5.8 GB | ≈ 27–41 tok/s | ≈ 39–58 tok/s | ≈ 62–93 tok/s | ≈ 70–105 tok/s |
| GLM-OCR 1.3BDocuments and OCR · Images | 5.8 GB | ≈ 22–33 tok/s | ≈ 31–47 tok/s | ≈ 51–76 tok/s | ≈ 57–85 tok/s |
| Qwen 3.5 4BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 5.3 GB | ≈ 18–27 tok/s | ≈ 26–39 tok/s | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen3 Embedding 0.6BSearch your documents (RAG) | 5.3 GB | ≈ 94–141 tok/s | ≈ 133–199 tok/s | 150+ tok/s | 150+ tok/s |
| Bonsai 2 9B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 5.2 GB | ≈ 19–28 tok/s | ≈ 27–40 tok/s | ≈ 43–64 tok/s | ≈ 48–72 tok/s |
| TranslateGemma 4BTranslation · Many languages | 5.1 GB | ≈ 18–27 tok/s | ≈ 26–39 tok/s | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen 3.5 2BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 3.3 GB | ≈ 32–47 tok/s | ≈ 45–67 tok/s | ≈ 72–108 tok/s | ≈ 81–121 tok/s |
| DeepSeek-R1 (distilled) 1.5BReasoning and math · Coding | 3.0 GB | ≈ 55–82 tok/s | ≈ 77–116 tok/s | ≈ 124–186 tok/s | ≈ 140–209 tok/s |
| Llama 3.2 1BChat and writing · Documents and OCR · Many languages | 2.9 GB | ≈ 74–111 tok/s | ≈ 105–157 tok/s | 150+ tok/s | 150+ tok/s |
| Qwen 3.5 0.8BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 2.7 GB | ≈ 46–69 tok/s | ≈ 65–98 tok/s | ≈ 105–158 tok/s | ≈ 118–177 tok/s |
| LFM3 1.7BAI agents · Chat and writing | 2.5 GB | ≈ 55–82 tok/s | ≈ 77–116 tok/s | ≈ 124–186 tok/s | ≈ 140–209 tok/s |
What can a 32 GB Mac mini run?
32 GB is where local LLMs get interesting: about 24 GB for the GPU opens the 24–35B generation, including the coding agents Qwen3-Coder, North Mini Code and Devstral Small 2, Qwen 3.8 27B, Gemma 4 26B and Muse Glimmer for agents. 32 GB exists only with the M4 and M6, whose bandwidth is lower, so dense 24–27B models run slowly; mixture-of-experts models with about 3B active parameters stay fast.
| Model | Memory needed | Speed on M4 | Speed on M6 |
|---|---|---|---|
| Qwen 3.6 35B (MoE, 3B active)Coding · AI agents · Reasoning and math · Chat and writing · Images | 26 GBtight | ≈ 30–45 tok/s | ≈ 43–64 tok/s |
| Flint 40B (MoE, 2.5B active)Coding · AI agents | 25 GBtight | ≈ 42–62 tok/s | ≈ 59–88 tok/s |
| Ornith 1.5 35B (MoE, 3B active)Coding · AI agents | 25 GBtight | ≈ 31–47 tok/s | ≈ 44–67 tok/s |
| Qwen3-VL 30B (MoE, 3B active)Images · Documents and OCR · AI agents | 24 GBtight | ≈ 31–47 tok/s | ≈ 44–66 tok/s |
| FLUX.1 [dev] 11.9BImage generation | from 24 GB Macreported | see reports | see reports |
| Qwen-Image 2.1 7.1BImage generation | from 24 GB Macreported | see reports | see reports |
| ACE-Step 1.5 2B+Music | from 24 GB Macreported | see reports | see reports |
| TRELLIS.2 4B3D models | from 24 GB Macreported | see reports | see reports |
| Hunyuan3D 2.1 3.3B3D models | from 24 GB Macreported | see reports | see reports |
| Qwen 3.5 35B (MoE, 3B active)Chat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 24 GB | ≈ 33–49 tok/s | ≈ 46–70 tok/s |
| Qwen3-Coder 30B (MoE, 3.3B active)Coding · AI agents | 23 GB | ≈ 29–44 tok/s | ≈ 41–62 tok/s |
| Mistral Small 3.2 24BChat and writing · Images · AI agents | 21 GB | ≈ 4–6 tok/s | ≈ 6–9 tok/s |
| Devstral Small 2 24BCoding · AI agents · Images | 21 GB | ≈ 4–6 tok/s | ≈ 6–9 tok/s |
| North Mini Code 1.0 30B (MoE, 3B active)Coding · AI agents | 21 GB | ≈ 32–48 tok/s | ≈ 45–68 tok/s |
| Qwen 3.8 27BCoding · AI agents · Reasoning and math · Images | 21 GB | ≈ 3–5 tok/s | ≈ 5–7 tok/s |
| TranslateGemma 27BTranslation · Many languages | 21 GB | ≈ 4–5 tok/s | ≈ 5–8 tok/s |
| Gemma 4 26B (MoE, 4B active)Chat and writing · Reasoning and math · AI agents · Coding · Images | 21 GB | ≈ 22–32 tok/s | ≈ 30–46 tok/s |
| Qwen 3.5 27BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 20 GB | ≈ 4–5 tok/s | ≈ 5–8 tok/s |
| Qwen 3.6 27BCoding · AI agents · Reasoning and math · Chat and writing · Images | 20 GB | ≈ 4–5 tok/s | ≈ 5–8 tok/s |
| Muse Glimmer 30BAI agents · Reasoning and math · Images | 20 GB | ≈ 3–5 tok/s | ≈ 5–7 tok/s |
| DeepSeek-R1 (distilled) 14BReasoning and math · Coding | 16 GB | ≈ 7–10 tok/s | ≈ 9–14 tok/s |
| FLUX.2 [klein] 3.9BImage generation | from 16 GB Macreported | see reports | see reports |
| Z-Image-Turbo 6.2BImage generation | from 16 GB Macreported | see reports | see reports |
| Krea 2 Turbo 12.8BImage generation | from 16 GB Macreported | see reports | see reports |
| Stable Diffusion XL 2.6BImage generation | from 16 GB Macreported | see reports | see reports |
| MiniMax H3 33.1BVideo generation | from 16 GB Macreported | see reports | see reports |
| LTX-2.5 22BVideo generation | from 16 GB Macreported | see reports | see reports |
| Whisper large-v3-turbo 809MSpeech to text | from 16 GB Macreported | see reports | see reports |
| Parakeet TDT 0.6B v3 627MSpeech to text | from 16 GB Macreported | see reports | see reports |
| Qwen3-ASR 2.3BSpeech to text | from 16 GB Macreported | see reports | see reports |
| Kokoro-82MText to speech | from 16 GB Macreported | see reports | see reports |
| Qwen3-TTS 1.9BText to speech | from 16 GB Macreported | see reports | see reports |
| Chatterbox 0.5BText to speech | from 16 GB Macreported | see reports | see reports |
| Phonon 41MText to speech | from 16 GB Macreported | see reports | see reports |
| gpt-oss 20B (MoE, 3.6B active)Reasoning and math · AI agents · Coding | 16 GB | ≈ 25–37 tok/s | ≈ 35–53 tok/s |
| Phi-4 14BReasoning and math · Chat and writing | 13 GB | ≈ 7–10 tok/s | ≈ 9–14 tok/s |
| Qwen3-VL 8BImages · Documents and OCR · AI agents | 12 GB | ≈ 10–15 tok/s | ≈ 14–21 tok/s |
| Granite 4.2 8BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 12 GB | ≈ 11–17 tok/s | ≈ 16–24 tok/s |
| TranslateGemma 12BTranslation · Many languages | 12 GB | ≈ 7–11 tok/s | ≈ 10–16 tok/s |
| Gemma 4 12BChat and writing · Reasoning and math · AI agents · Coding · Images | 11 GB | ≈ 8–11 tok/s | ≈ 11–16 tok/s |
| Bonsai 2 27B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 11 GB | ≈ 7–11 tok/s | ≈ 10–15 tok/s |
| Gemma 3n E4BChat and writing · Images · Many languages | 11 GB | ≈ 6–9 tok/s | ≈ 9–13 tok/s |
| DeepSeek-R1 (distilled) 8B (0528)Reasoning and math · Coding | 11 GB | ≈ 12–17 tok/s | ≈ 16–25 tok/s |
| Qwen3 Embedding 8BSearch your documents (RAG) | 10 GB | ≈ 13–19 tok/s | ≈ 18–27 tok/s |
| Llama 3.1 8BChat and writing · Many languages · Documents and OCR · Coding | 10 GB | ≈ 12–18 tok/s | ≈ 17–26 tok/s |
| Nemotron 3.5 Bolt E4BChat and writing · Coding · AI agents · Images | 9.4 GB | ≈ 8–11 tok/s | ≈ 11–16 tok/s |
| Gemma 3n E2BChat and writing · Images · Many languages | 9.2 GB | ≈ 8–12 tok/s | ≈ 11–16 tok/s |
| Qwen3-VL 4BImages · Documents and OCR · AI agents | 9.0 GB | ≈ 18–27 tok/s | ≈ 26–39 tok/s |
| Qwen 3.5 9BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 8.6 GB | ≈ 9–14 tok/s | ≈ 13–19 tok/s |
| Ornith 1.5 9BCoding · AI agents | 8.6 GB | ≈ 9–14 tok/s | ≈ 13–19 tok/s |
| Qwen3 Embedding 4BSearch your documents (RAG) | 8.2 GB | ≈ 24–36 tok/s | ≈ 34–51 tok/s |
| Gemma 4 E4BChat and writing · Reasoning and math · AI agents · Coding · Images | 8.1 GB | ≈ 9–14 tok/s | ≈ 13–19 tok/s |
| DeepSeek-R1 (distilled) 7BReasoning and math · Coding | 7.5 GB | ≈ 13–19 tok/s | ≈ 18–27 tok/s |
| Llama 3.2 3BChat and writing · Documents and OCR · Many languages | 6.7 GB | ≈ 30–45 tok/s | ≈ 43–64 tok/s |
| Qwen3-VL 2BImages · Documents and OCR · AI agents | 6.6 GB | ≈ 32–47 tok/s | ≈ 45–67 tok/s |
| Gemma 4 E2BChat and writing · Reasoning and math · AI agents · Coding · Images | 5.8 GB | ≈ 13–20 tok/s | ≈ 18–28 tok/s |
| Granite 4.2 3BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 5.8 GB | ≈ 27–41 tok/s | ≈ 39–58 tok/s |
| GLM-OCR 1.3BDocuments and OCR · Images | 5.8 GB | ≈ 22–33 tok/s | ≈ 31–47 tok/s |
| Qwen 3.5 4BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 5.3 GB | ≈ 18–27 tok/s | ≈ 26–39 tok/s |
| Qwen3 Embedding 0.6BSearch your documents (RAG) | 5.3 GB | ≈ 94–141 tok/s | ≈ 133–199 tok/s |
| Bonsai 2 9B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 5.2 GB | ≈ 19–28 tok/s | ≈ 27–40 tok/s |
| TranslateGemma 4BTranslation · Many languages | 5.1 GB | ≈ 18–27 tok/s | ≈ 26–39 tok/s |
| Qwen 3.5 2BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 3.3 GB | ≈ 32–47 tok/s | ≈ 45–67 tok/s |
| DeepSeek-R1 (distilled) 1.5BReasoning and math · Coding | 3.0 GB | ≈ 55–82 tok/s | ≈ 77–116 tok/s |
| Llama 3.2 1BChat and writing · Documents and OCR · Many languages | 2.9 GB | ≈ 74–111 tok/s | ≈ 105–157 tok/s |
| Qwen 3.5 0.8BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 2.7 GB | ≈ 46–69 tok/s | ≈ 65–98 tok/s |
| LFM3 1.7BAI agents · Chat and writing | 2.5 GB | ≈ 55–82 tok/s | ≈ 77–116 tok/s |
What does a 48 GB Mac mini add?
48 GB (M4 Pro or M5 Pro) gives about 36 GB to the GPU: it adds the dense 30–32B models (Gemma 4 31B, Qwen3-VL 32B, DeepSeek-R1 32B, Granite 4.2 30B) and NVIDIA's Nemotron 3.5 Lightning for agents, and it runs everything from the 32 GB list with room for longer contexts. The Pro chips' bandwidth also makes every model roughly twice as fast as on an M4.
| Model | Memory needed | Speed on M4 Pro | Speed on M5 Pro |
|---|---|---|---|
| Qwen3-VL 32BImages · Documents and OCR · AI agents | 30 GB | ≈ 7–10 tok/s | ≈ 7–11 tok/s |
| DeepSeek-R1 (distilled) 32BReasoning and math · Coding | 29 GB | ≈ 7–10 tok/s | ≈ 8–12 tok/s |
| Nemotron 3.5 Lightning 30B (MoE, 3B active)AI agents · Chat and writing | 28 GB | ≈ 58–86 tok/s | ≈ 65–97 tok/s |
| Granite 4.2 30BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 27 GB | ≈ 8–11 tok/s | ≈ 9–13 tok/s |
| Gemma 4 31BChat and writing · Reasoning and math · AI agents · Coding · Images | 27 GB | ≈ 7–10 tok/s | ≈ 8–12 tok/s |
| Qwen 3.6 35B (MoE, 3B active)Coding · AI agents · Reasoning and math · Chat and writing · Images | 26 GB | ≈ 68–102 tok/s | ≈ 77–115 tok/s |
| Flint 40B (MoE, 2.5B active)Coding · AI agents | 25 GB | ≈ 94–142 tok/s | ≈ 106–159 tok/s |
| Ornith 1.5 35B (MoE, 3B active)Coding · AI agents | 25 GB | ≈ 71–107 tok/s | ≈ 80–120 tok/s |
| Qwen3-VL 30B (MoE, 3B active)Images · Documents and OCR · AI agents | 24 GB | ≈ 71–106 tok/s | ≈ 80–119 tok/s |
| FLUX.1 [dev] 11.9BImage generation | from 24 GB Macreported | see reports | see reports |
| Qwen-Image 2.1 7.1BImage generation | from 24 GB Macreported | see reports | see reports |
| ACE-Step 1.5 2B+Music | from 24 GB Macreported | see reports | see reports |
| TRELLIS.2 4B3D models | from 24 GB Macreported | see reports | see reports |
| Hunyuan3D 2.1 3.3B3D models | from 24 GB Macreported | see reports | see reports |
| Qwen 3.5 35B (MoE, 3B active)Chat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 24 GB | ≈ 74–112 tok/s | ≈ 84–126 tok/s |
| Qwen3-Coder 30B (MoE, 3.3B active)Coding · AI agents | 23 GB | ≈ 66–100 tok/s | ≈ 75–112 tok/s |
| Mistral Small 3.2 24BChat and writing · Images · AI agents | 21 GB | ≈ 9–14 tok/s | ≈ 10–15 tok/s |
| Devstral Small 2 24BCoding · AI agents · Images | 21 GB | ≈ 9–14 tok/s | ≈ 10–15 tok/s |
| North Mini Code 1.0 30B (MoE, 3B active)Coding · AI agents | 21 GB | ≈ 73–110 tok/s | ≈ 82–123 tok/s |
| Qwen 3.8 27BCoding · AI agents · Reasoning and math · Images | 21 GB | ≈ 8–11 tok/s | ≈ 9–13 tok/s |
| TranslateGemma 27BTranslation · Many languages | 21 GB | ≈ 8–12 tok/s | ≈ 9–14 tok/s |
| Gemma 4 26B (MoE, 4B active)Chat and writing · Reasoning and math · AI agents · Coding · Images | 21 GB | ≈ 49–73 tok/s | ≈ 55–83 tok/s |
| Qwen 3.5 27BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 20 GB | ≈ 8–12 tok/s | ≈ 9–14 tok/s |
| Qwen 3.6 27BCoding · AI agents · Reasoning and math · Chat and writing · Images | 20 GB | ≈ 8–12 tok/s | ≈ 9–14 tok/s |
| Muse Glimmer 30BAI agents · Reasoning and math · Images | 20 GB | ≈ 8–11 tok/s | ≈ 9–13 tok/s |
| DeepSeek-R1 (distilled) 14BReasoning and math · Coding | 16 GB | ≈ 15–23 tok/s | ≈ 17–26 tok/s |
| FLUX.2 [klein] 3.9BImage generation | from 16 GB Macreported | see reports | see reports |
| Z-Image-Turbo 6.2BImage generation | from 16 GB Macreported | see reports | see reports |
| Krea 2 Turbo 12.8BImage generation | from 16 GB Macreported | see reports | see reports |
| Stable Diffusion XL 2.6BImage generation | from 16 GB Macreported | see reports | see reports |
| MiniMax H3 33.1BVideo generation | from 16 GB Macreported | see reports | see reports |
| LTX-2.5 22BVideo generation | from 16 GB Macreported | see reports | see reports |
| Whisper large-v3-turbo 809MSpeech to text | from 16 GB Macreported | see reports | see reports |
| Parakeet TDT 0.6B v3 627MSpeech to text | from 16 GB Macreported | see reports | see reports |
| Qwen3-ASR 2.3BSpeech to text | from 16 GB Macreported | see reports | see reports |
| Kokoro-82MText to speech | from 16 GB Macreported | see reports | see reports |
| Qwen3-TTS 1.9BText to speech | from 16 GB Macreported | see reports | see reports |
| Chatterbox 0.5BText to speech | from 16 GB Macreported | see reports | see reports |
| Phonon 41MText to speech | from 16 GB Macreported | see reports | see reports |
| gpt-oss 20B (MoE, 3.6B active)Reasoning and math · AI agents · Coding | 16 GB | ≈ 57–85 tok/s | ≈ 64–95 tok/s |
| Phi-4 14BReasoning and math · Chat and writing | 13 GB | ≈ 15–23 tok/s | ≈ 17–25 tok/s |
| Qwen3-VL 8BImages · Documents and OCR · AI agents | 12 GB | ≈ 22–33 tok/s | ≈ 25–37 tok/s |
| Granite 4.2 8BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 12 GB | ≈ 26–39 tok/s | ≈ 29–43 tok/s |
| TranslateGemma 12BTranslation · Many languages | 12 GB | ≈ 17–25 tok/s | ≈ 19–28 tok/s |
| Gemma 4 12BChat and writing · Reasoning and math · AI agents · Coding · Images | 11 GB | ≈ 17–26 tok/s | ≈ 19–29 tok/s |
| Bonsai 2 27B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 11 GB | ≈ 16–25 tok/s | ≈ 18–28 tok/s |
| Gemma 3n E4BChat and writing · Images · Many languages | 11 GB | ≈ 14–21 tok/s | ≈ 16–23 tok/s |
| DeepSeek-R1 (distilled) 8B (0528)Reasoning and math · Coding | 11 GB | ≈ 26–39 tok/s | ≈ 30–44 tok/s |
| Qwen3 Embedding 8BSearch your documents (RAG) | 10 GB | ≈ 29–44 tok/s | ≈ 33–49 tok/s |
| Llama 3.1 8BChat and writing · Many languages · Documents and OCR · Coding | 10 GB | ≈ 28–42 tok/s | ≈ 31–47 tok/s |
| Nemotron 3.5 Bolt E4BChat and writing · Coding · AI agents · Images | 9.4 GB | ≈ 17–26 tok/s | ≈ 19–29 tok/s |
| Gemma 3n E2BChat and writing · Images · Many languages | 9.2 GB | ≈ 18–26 tok/s | ≈ 20–30 tok/s |
| Qwen3-VL 4BImages · Documents and OCR · AI agents | 9.0 GB | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen 3.5 9BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 8.6 GB | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| Ornith 1.5 9BCoding · AI agents | 8.6 GB | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| Qwen3 Embedding 4BSearch your documents (RAG) | 8.2 GB | ≈ 55–82 tok/s | ≈ 61–92 tok/s |
| Gemma 4 E4BChat and writing · Reasoning and math · AI agents · Coding · Images | 8.1 GB | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| DeepSeek-R1 (distilled) 7BReasoning and math · Coding | 7.5 GB | ≈ 29–44 tok/s | ≈ 33–49 tok/s |
| Llama 3.2 3BChat and writing · Documents and OCR · Many languages | 6.7 GB | ≈ 68–102 tok/s | ≈ 77–115 tok/s |
| Qwen3-VL 2BImages · Documents and OCR · AI agents | 6.6 GB | ≈ 72–108 tok/s | ≈ 81–121 tok/s |
| Gemma 4 E2BChat and writing · Reasoning and math · AI agents · Coding · Images | 5.8 GB | ≈ 30–45 tok/s | ≈ 33–50 tok/s |
| Granite 4.2 3BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 5.8 GB | ≈ 62–93 tok/s | ≈ 70–105 tok/s |
| GLM-OCR 1.3BDocuments and OCR · Images | 5.8 GB | ≈ 51–76 tok/s | ≈ 57–85 tok/s |
| Qwen 3.5 4BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 5.3 GB | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen3 Embedding 0.6BSearch your documents (RAG) | 5.3 GB | 150+ tok/s | 150+ tok/s |
| Bonsai 2 9B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 5.2 GB | ≈ 43–64 tok/s | ≈ 48–72 tok/s |
| TranslateGemma 4BTranslation · Many languages | 5.1 GB | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen 3.5 2BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 3.3 GB | ≈ 72–108 tok/s | ≈ 81–121 tok/s |
| DeepSeek-R1 (distilled) 1.5BReasoning and math · Coding | 3.0 GB | ≈ 124–186 tok/s | ≈ 140–209 tok/s |
| Llama 3.2 1BChat and writing · Documents and OCR · Many languages | 2.9 GB | 150+ tok/s | 150+ tok/s |
| Qwen 3.5 0.8BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 2.7 GB | ≈ 105–158 tok/s | ≈ 118–177 tok/s |
| LFM3 1.7BAI agents · Chat and writing | 2.5 GB | ≈ 124–186 tok/s | ≈ 140–209 tok/s |
Can a 64 GB Mac mini run a 70B model?
Only tightly. At 4-bit with a 32K context, Llama 3.3 70B and the DeepSeek-R1 70B distill need about 54 GB, more than the default 48 GB GPU share, so you must raise the limit (sudo sysctl iogpu.wired_limit_mb=58000, resets at restart), and expect a few tokens per second. Otherwise 64 GB runs the same models as 48 GB; the extra memory is for long contexts, full 8-bit builds or two models at once (for example a chat model and an embedding model for search over documents).
| Model | Memory needed | Speed on M4 Pro | Speed on M5 Pro |
|---|---|---|---|
| Wan 2.2 (14B) 14.3BVideo generation | from 64 GB Macreported | see reports | see reports |
| Llama 3.3 70BChat and writing · Many languages | 54 GBtight | ≈ 3–5 tok/s | ≈ 4–5 tok/s |
| DeepSeek-R1 (distilled) 70BReasoning and math · Coding | 54 GBtight | ≈ 3–5 tok/s | ≈ 4–5 tok/s |
| Qwen3-VL 32BImages · Documents and OCR · AI agents | 30 GB | ≈ 7–10 tok/s | ≈ 7–11 tok/s |
| DeepSeek-R1 (distilled) 32BReasoning and math · Coding | 29 GB | ≈ 7–10 tok/s | ≈ 8–12 tok/s |
| Nemotron 3.5 Lightning 30B (MoE, 3B active)AI agents · Chat and writing | 28 GB | ≈ 58–86 tok/s | ≈ 65–97 tok/s |
| Granite 4.2 30BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 27 GB | ≈ 8–11 tok/s | ≈ 9–13 tok/s |
| Gemma 4 31BChat and writing · Reasoning and math · AI agents · Coding · Images | 27 GB | ≈ 7–10 tok/s | ≈ 8–12 tok/s |
| Qwen 3.6 35B (MoE, 3B active)Coding · AI agents · Reasoning and math · Chat and writing · Images | 26 GB | ≈ 68–102 tok/s | ≈ 77–115 tok/s |
| Flint 40B (MoE, 2.5B active)Coding · AI agents | 25 GB | ≈ 94–142 tok/s | ≈ 106–159 tok/s |
| Ornith 1.5 35B (MoE, 3B active)Coding · AI agents | 25 GB | ≈ 71–107 tok/s | ≈ 80–120 tok/s |
| Qwen3-VL 30B (MoE, 3B active)Images · Documents and OCR · AI agents | 24 GB | ≈ 71–106 tok/s | ≈ 80–119 tok/s |
| FLUX.1 [dev] 11.9BImage generation | from 24 GB Macreported | see reports | see reports |
| Qwen-Image 2.1 7.1BImage generation | from 24 GB Macreported | see reports | see reports |
| ACE-Step 1.5 2B+Music | from 24 GB Macreported | see reports | see reports |
| TRELLIS.2 4B3D models | from 24 GB Macreported | see reports | see reports |
| Hunyuan3D 2.1 3.3B3D models | from 24 GB Macreported | see reports | see reports |
| Qwen 3.5 35B (MoE, 3B active)Chat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 24 GB | ≈ 74–112 tok/s | ≈ 84–126 tok/s |
| Qwen3-Coder 30B (MoE, 3.3B active)Coding · AI agents | 23 GB | ≈ 66–100 tok/s | ≈ 75–112 tok/s |
| Mistral Small 3.2 24BChat and writing · Images · AI agents | 21 GB | ≈ 9–14 tok/s | ≈ 10–15 tok/s |
| Devstral Small 2 24BCoding · AI agents · Images | 21 GB | ≈ 9–14 tok/s | ≈ 10–15 tok/s |
| North Mini Code 1.0 30B (MoE, 3B active)Coding · AI agents | 21 GB | ≈ 73–110 tok/s | ≈ 82–123 tok/s |
| Qwen 3.8 27BCoding · AI agents · Reasoning and math · Images | 21 GB | ≈ 8–11 tok/s | ≈ 9–13 tok/s |
| TranslateGemma 27BTranslation · Many languages | 21 GB | ≈ 8–12 tok/s | ≈ 9–14 tok/s |
| Gemma 4 26B (MoE, 4B active)Chat and writing · Reasoning and math · AI agents · Coding · Images | 21 GB | ≈ 49–73 tok/s | ≈ 55–83 tok/s |
| Qwen 3.5 27BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 20 GB | ≈ 8–12 tok/s | ≈ 9–14 tok/s |
| Qwen 3.6 27BCoding · AI agents · Reasoning and math · Chat and writing · Images | 20 GB | ≈ 8–12 tok/s | ≈ 9–14 tok/s |
| Muse Glimmer 30BAI agents · Reasoning and math · Images | 20 GB | ≈ 8–11 tok/s | ≈ 9–13 tok/s |
| DeepSeek-R1 (distilled) 14BReasoning and math · Coding | 16 GB | ≈ 15–23 tok/s | ≈ 17–26 tok/s |
| FLUX.2 [klein] 3.9BImage generation | from 16 GB Macreported | see reports | see reports |
| Z-Image-Turbo 6.2BImage generation | from 16 GB Macreported | see reports | see reports |
| Krea 2 Turbo 12.8BImage generation | from 16 GB Macreported | see reports | see reports |
| Stable Diffusion XL 2.6BImage generation | from 16 GB Macreported | see reports | see reports |
| MiniMax H3 33.1BVideo generation | from 16 GB Macreported | see reports | see reports |
| LTX-2.5 22BVideo generation | from 16 GB Macreported | see reports | see reports |
| Whisper large-v3-turbo 809MSpeech to text | from 16 GB Macreported | see reports | see reports |
| Parakeet TDT 0.6B v3 627MSpeech to text | from 16 GB Macreported | see reports | see reports |
| Qwen3-ASR 2.3BSpeech to text | from 16 GB Macreported | see reports | see reports |
| Kokoro-82MText to speech | from 16 GB Macreported | see reports | see reports |
| Qwen3-TTS 1.9BText to speech | from 16 GB Macreported | see reports | see reports |
| Chatterbox 0.5BText to speech | from 16 GB Macreported | see reports | see reports |
| Phonon 41MText to speech | from 16 GB Macreported | see reports | see reports |
| gpt-oss 20B (MoE, 3.6B active)Reasoning and math · AI agents · Coding | 16 GB | ≈ 57–85 tok/s | ≈ 64–95 tok/s |
| Phi-4 14BReasoning and math · Chat and writing | 13 GB | ≈ 15–23 tok/s | ≈ 17–25 tok/s |
| Qwen3-VL 8BImages · Documents and OCR · AI agents | 12 GB | ≈ 22–33 tok/s | ≈ 25–37 tok/s |
| Granite 4.2 8BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 12 GB | ≈ 26–39 tok/s | ≈ 29–43 tok/s |
| TranslateGemma 12BTranslation · Many languages | 12 GB | ≈ 17–25 tok/s | ≈ 19–28 tok/s |
| Gemma 4 12BChat and writing · Reasoning and math · AI agents · Coding · Images | 11 GB | ≈ 17–26 tok/s | ≈ 19–29 tok/s |
| Bonsai 2 27B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 11 GB | ≈ 16–25 tok/s | ≈ 18–28 tok/s |
| Gemma 3n E4BChat and writing · Images · Many languages | 11 GB | ≈ 14–21 tok/s | ≈ 16–23 tok/s |
| DeepSeek-R1 (distilled) 8B (0528)Reasoning and math · Coding | 11 GB | ≈ 26–39 tok/s | ≈ 30–44 tok/s |
| Qwen3 Embedding 8BSearch your documents (RAG) | 10 GB | ≈ 29–44 tok/s | ≈ 33–49 tok/s |
| Llama 3.1 8BChat and writing · Many languages · Documents and OCR · Coding | 10 GB | ≈ 28–42 tok/s | ≈ 31–47 tok/s |
| Nemotron 3.5 Bolt E4BChat and writing · Coding · AI agents · Images | 9.4 GB | ≈ 17–26 tok/s | ≈ 19–29 tok/s |
| Gemma 3n E2BChat and writing · Images · Many languages | 9.2 GB | ≈ 18–26 tok/s | ≈ 20–30 tok/s |
| Qwen3-VL 4BImages · Documents and OCR · AI agents | 9.0 GB | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen 3.5 9BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 8.6 GB | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| Ornith 1.5 9BCoding · AI agents | 8.6 GB | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| Qwen3 Embedding 4BSearch your documents (RAG) | 8.2 GB | ≈ 55–82 tok/s | ≈ 61–92 tok/s |
| Gemma 4 E4BChat and writing · Reasoning and math · AI agents · Coding · Images | 8.1 GB | ≈ 21–31 tok/s | ≈ 23–35 tok/s |
| DeepSeek-R1 (distilled) 7BReasoning and math · Coding | 7.5 GB | ≈ 29–44 tok/s | ≈ 33–49 tok/s |
| Llama 3.2 3BChat and writing · Documents and OCR · Many languages | 6.7 GB | ≈ 68–102 tok/s | ≈ 77–115 tok/s |
| Qwen3-VL 2BImages · Documents and OCR · AI agents | 6.6 GB | ≈ 72–108 tok/s | ≈ 81–121 tok/s |
| Gemma 4 E2BChat and writing · Reasoning and math · AI agents · Coding · Images | 5.8 GB | ≈ 30–45 tok/s | ≈ 33–50 tok/s |
| Granite 4.2 3BChat and writing · Many languages · Coding · AI agents · Search your documents (RAG) | 5.8 GB | ≈ 62–93 tok/s | ≈ 70–105 tok/s |
| GLM-OCR 1.3BDocuments and OCR · Images | 5.8 GB | ≈ 51–76 tok/s | ≈ 57–85 tok/s |
| Qwen 3.5 4BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 5.3 GB | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen3 Embedding 0.6BSearch your documents (RAG) | 5.3 GB | 150+ tok/s | 150+ tok/s |
| Bonsai 2 9B (ternary)Chat and writing · Reasoning and math · AI agents · Images | 5.2 GB | ≈ 43–64 tok/s | ≈ 48–72 tok/s |
| TranslateGemma 4BTranslation · Many languages | 5.1 GB | ≈ 41–62 tok/s | ≈ 47–70 tok/s |
| Qwen 3.5 2BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 3.3 GB | ≈ 72–108 tok/s | ≈ 81–121 tok/s |
| DeepSeek-R1 (distilled) 1.5BReasoning and math · Coding | 3.0 GB | ≈ 124–186 tok/s | ≈ 140–209 tok/s |
| Llama 3.2 1BChat and writing · Documents and OCR · Many languages | 2.9 GB | 150+ tok/s | 150+ tok/s |
| Qwen 3.5 0.8BChat and writing · Reasoning and math · Coding · AI agents · Images · Many languages | 2.7 GB | ≈ 105–158 tok/s | ≈ 118–177 tok/s |
| LFM3 1.7BAI agents · Chat and writing | 2.5 GB | ≈ 124–186 tok/s | ≈ 140–209 tok/s |
Rent or buy the Mac mini for local LLMs?
Renting lets you test a model on exactly the chip and memory you are considering, or run an always-on agent without hardware at home. For agent hosting see Mac mini for OpenClaw and AI agents: rent, buy or VPS?; for 48 and 64 GB plans see Rent a Mac mini M4 Pro or M5 Pro in the cloud; and every plan we track is in the price tables for Europe, North America and Asia and Australia.
Frequently asked questions
What LLM can a Mac mini M4 with 16 GB run?
Models up to about 8 to 12 billion parameters at 4-bit, such as Gemma 4 12B, Qwen 3.5 9B, Llama 3.1 8B, the DeepSeek-R1 8B distill, Ornith 1.5 9B for coding, Qwen3-VL 8B for images and GLM-OCR for documents. macOS lets the GPU use about 12 GB of the 16 GB by default, so 20B-class models do not fit at 4-bit. The exception is Bonsai 2 27B: its developer stores the weights in ternary form (-1, 0 or +1), so the file is about 8 GB.
Is 24 GB enough for local LLMs on a Mac mini?
It adds only a few models over 16 GB: gpt-oss 20B, the DeepSeek-R1 14B distill and Phi-4 (14B). The popular 24–35B models need about 20–24 GB of memory with a 32K context, which is more than the roughly 18 GB macOS gives the GPU on a 24 GB Mac. If you want those models, go to 32 GB.
What can a 32 GB Mac mini run?
The 24–35B models that fit in about 24 GB: Qwen3-Coder 30B, North Mini Code, Devstral Small 2, Qwen 3.8 27B, Qwen 3.5 27B and 35B, Gemma 4 26B, Mistral Small 3.2 and Muse Glimmer. A 32 GB Mac mini comes with the M4 or M6; mixture-of-experts models such as Qwen3-Coder run fast even on their lower memory bandwidth.
Can a 64 GB Mac mini run a 70B model?
Only tightly. Llama 3.3 70B and the DeepSeek-R1 70B distill need about 54 GB at 4-bit with a 32K context, more than the default GPU share of 64 GB, so you must raise the GPU memory limit, and speed is a few tokens per second. Over 48 GB, 64 GB mostly buys room for longer contexts or two models at once.
Is a Mac mini M4 Pro faster for LLMs than an M4?
Yes, for the same model: Apple lists 273 GB/s of memory bandwidth for the M4 Pro against 120 GB/s for the M4, and generating text is limited mostly by bandwidth, so the M4 Pro writes roughly twice as fast. The M5 Pro (307 GB/s) is faster still. But the M4 Pro starts at 24 GB and has no 32 GB option.