Best local AI models for the M4 Pro with 48 GB
Of its 48 GB of unified memory, macOS lets the GPU use about 75%, so roughly 36 GB is left for a model and its context. That fits 2 790 open-weight models. Speeds come from the chip’s 273 GB/s of memory bandwidth.
The M4 Pro (48 GB) has 48 GB of unified memory shared between the CPU and GPU, with about 36 GB usable for models. Memory bandwidth is 273 GB/s, which sets how quickly a model can generate tokens once it is loaded. Models that fit within that 36 GB budget include Qwen3.8-27B, gemma-4-31B-it, and Qwen3.6-35B-A3B. Smaller options such as Llama-3.1-8B-Instruct also run comfortably on this machine.
- chip
- M4 Pro
- unified memory
- 48 GB
- usable by the GPU
- ≈36 GB
- bandwidth
- 273 GB/s
Chat & Text Generation
All 1 080 that fit →| # | Model | Params | Best quant | Memory | Speed |
|---|---|---|---|---|---|
| 1 | 36BA3B | Q6_K | ≈31 GBFits in memory | ~61 tok/s | |
| 2 | 39.8BA1.3B | Q6_K | ≈33 GBFits in memory | ~146 tok/s | |
| 3 | 35.1BA3.9B | Q6_K | ≈29 GBFits in memory | ~47 tok/s | |
| 4 | 36.2B | Q6_K | ≈32 GBFits in memory | ~5 tok/s | |
| 5 | 31.2BA4.6B | Q6_K | ≈32 GBFits in memory | ~40 tok/s | |
| 6 | 31.2BA4B | Q5_K_M | ≈31 GBFits in memory | ~53 tok/s | |
| 7 | 33.1B | Q6_K | ≈31 GBFits in memory | ~5.5 tok/s | |
| 8 | 36BA4.7B | Q6_K | ≈30 GBFits in memory | ~38 tok/s |
Code Models
All 107 that fit →| # | Model | Params | Best quant | Memory | Speed |
|---|---|---|---|---|---|
| 1 | 12.1BA2.5B | Q8_0 | ≈14 GBFits in memory | ~57 tok/s | |
| 2 | 32.8B | Q6_K | ≈29 GBFits in memory | ~5.6 tok/s | |
| 3 | 34.7BA3.5B | Q6_K | ≈31 GBFits in memory | ~53 tok/s | |
| 4 | 30.5BA3B | Q8_0 | ≈33 GBFits in memory | ~47 tok/s | |
| 5 | 30.5BA3.3B | Q8_0 | ≈33 GBFits in memory | ~43 tok/s | |
| 6 | 34B | Q6_K | ≈29 GBFits in memory | ~5.4 tok/s | |
| 7 | 39.8B | Q5_K_M | ≈31 GBFits in memory | ~5.3 tok/s | |
| 8 | 39.8B | Q5_K_M | ≈31 GBFits in memory | ~5.3 tok/s |
Reasoning Models
All 70 that fit →| # | Model | Params | Best quant | Memory | Speed |
|---|---|---|---|---|---|
| 1 | 32.8B | Q6_K | ≈29 GBFits in memory | ~5.6 tok/s | |
| 2 | 32.2B | Q6_K | ≈28 GBFits in memory | ~5.7 tok/s | |
| 3 | 32.8B | Q6_K | ≈29 GBFits in memory | ~5.6 tok/s | |
| 4 | 33.2B | Q6_K | ≈29 GBFits in memory | ~5.5 tok/s | |
| 5 | 32.1BA3B | Q6_K | ≈29 GBFits in memory | ~61 tok/s | |
| 6 | 33.2B | Q6_K | ≈29 GBFits in memory | ~5.5 tok/s | |
| 7 | 14.8B | Q8_0 | ≈17 GBFits in memory | ~9.6 tok/s | |
| 8 | 32.1BA3B | Q6_K | ≈29 GBFits in memory | ~61 tok/s |
Vision-Language Models
All 413 that fit →| # | Model | Params | Best quant | Memory | Speed |
|---|---|---|---|---|---|
| 1 | 27.8B | Q8_0 | ≈31 GBFits in memory | ~5.1 tok/s | |
| 2 | 31.3B | Q6_K | ≈33 GBFits in memory | ~5.8 tok/s | |
| 3 | 36BA3B | Q6_K | ≈30 GBFits in memory | ~61 tok/s | |
| 4 | 27.8B | Q8_0 | ≈30 GBFits in memory | ~5.1 tok/s | |
| 5 | 25.8BA4B | Q8_0 | ≈29 GBFits in memory | ~35 tok/s | |
| 6 | 29.8B | Q8_0 | ≈30 GBFits in memory | ~4.7 tok/s | |
| 7 | 36BA3B | Q6_K | ≈29 GBFits in memory | ~61 tok/s | |
| 8 | 35.2BA3B | Q6_K | ≈29 GBFits in memory | ~61 tok/s |
Text-to-Image
All 122 that fit →| # | Model | Params | Best quant | Memory |
|---|---|---|---|---|
| 1 | 7.1B | FP16 / BF16 | ≈16 GBFits in memory | |
| 2 | 7.1B | FP16 / BF16 | ≈16 GBFits in memory | |
| 3 | 12.8B | FP16 / BF16 | ≈29 GBFits in memory | |
| 4 | 20.4B | Q8_0 | ≈25 GBFits in memory | |
| 5 | 6.2B | FP16 / BF16 | ≈14 GBFits in memory | |
| 6 | 5.1B | FP16 / BF16 | ≈12 GBFits in memory | |
| 7 | 12.8B | FP16 / BF16 | ≈29 GBFits in memory | |
| 8 | 20.4B | Q8_0 | ≈25 GBFits in memory |
Speech-to-Text
All 100 that fit →| # | Model | Params | Best quant | Memory |
|---|---|---|---|---|
| 1 | 2.1B | FP16 / BF16 | ≈5.1 GBFits in memory | |
| 2 | 2.3B | FP16 / BF16 | ≈5.8 GBFits in memory | |
| 3 | 627M | FP16 / BF16 | ≈1.9 GBFits in memory | |
| 4 | 8.7B | FP16 / BF16 | ≈20 GBFits in memory | |
| 5 | 8.7B | FP16 / BF16 | ≈20 GBFits in memory | |
| 6 | 473M | FP16 / BF16 | ≈1.6 GBFits in memory | |
| 7 | 2.3B | FP16 / BF16 | ≈5.7 GBFits in memory | |
| 8 | 4.4B | FP16 / BF16 | ≈10 GBFits in memory |
Text-to-Speech
All 113 that fit →| # | Model | Params | Best quant | Memory |
|---|---|---|---|---|
| 1 | 3.6B | FP16 / BF16 | ≈8.6 GBFits in memory | |
| 2 | 3.5B | FP16 / BF16 | ≈8.3 GBFits in memory | |
| 3 | 2.3B | FP16 / BF16 | ≈5.6 GBFits in memory | |
| 4 | 613M | FP16 / BF16 | ≈1.9 GBFits in memory | |
| 5 | 1.9B | FP16 / BF16 | ≈4.8 GBFits in memory | |
| 6 | 1.5B | FP16 / BF16 | ≈3.8 GBFits in memory | |
| 7 | 915M | FP16 / BF16 | ≈2.5 GBFits in memory | |
| 8 | 4.6B | FP16 / BF16 | ≈11 GBFits in memory |
Embedding Models
All 244 that fit →| # | Model | Params | Best quant | Memory |
|---|---|---|---|---|
| 1 | 744M | FP16 / BF16 | ≈2.2 GBFits in memory | |
| 2 | 303M | FP16 / BF16 | ≈1.2 GBFits in memory | |
| 3 | 7.6B | FP16 / BF16 | ≈17 GBFits in memory | |
| 4 | 596M | FP16 / BF16 | ≈1.8 GBFits in memory | |
| 5 | 4B | FP16 / BF16 | ≈9.5 GBFits in memory | |
| 6 | 8B | FP16 / BF16 | ≈18 GBFits in memory | |
| 7 | 1.1B | FP16 / BF16 | ≈3.0 GBFits in memory | |
| 8 | 558M | FP16 / BF16 | ≈1.8 GBFits in memory |
Run it: MLX or llama.cpp (Metal)
Commands for Ornith-1.5-35B-A3B, the top chat pick here. MLX is Apple’s framework and usually generates a little faster; llama.cpp (also inside Ollama and LM Studio) runs any GGUF. Both use the GPU.
pip install mlx-lm
mlx_lm.generate --model mlx-community/Ornith-1.5-35B-A3B-OptiQ-4bit --prompt "Hello"Use `mlx_lm.chat` with the same --model for an interactive chat.
mlx_lm.server --model mlx-community/Ornith-1.5-35B-A3B-OptiQ-4bit --port 8080API on http://localhost:8080/v1.
ollama run ornith-1.5:35bThe library build picks its own default quantisation.
ollama run hf.co/bartowski/Ornith-1.5-35B-A3B-GGUF:Q6_K llama-server -hf bartowski/Ornith-1.5-35B-A3B-GGUF:Q6_Kbrew install llama.cpp ships with Metal. Downloads the GGUF on first run.
lms get bartowski/Ornith-1.5-35B-A3B-GGUF@Q6_KThen load it from the LM Studio app or with `lms load`.
Other Macs with similar memory
Frequently asked questions
- What is the best LLM for a M4 Pro with 48 GB?
- Our top chat pick is Ornith-1.5-35B-A3B (ornith-ai) at Q6_K, needing about 30.98 GB, at roughly 61 tokens/s.
- How much of the 48 GB can the GPU use?
- macOS lets the GPU wire about 75% of unified memory by default, so plan on roughly 36 GB for models. On recent macOS you can raise the limit with sudo sysctl iogpu.wired_limit_mb=<MB>; leave several GB for the system. The setting resets on reboot.
- MLX or llama.cpp on a Mac?
- MLX is Apple’s own framework and is usually a little faster for generation on Apple Silicon (mlx-community publishes ready-made 4-bit and 8-bit builds). llama.cpp with Metal, which Ollama and LM Studio use, has the widest model and GGUF coverage. Both run fully on the GPU.
- How fast is a M4 Pro with 48 GB for local LLMs?
- The M4 Pro has about 273 GB/s of memory bandwidth, which sets the ceiling for generation speed. Our estimate for Ornith-1.5-35B-A3B is about 61 tokens/s. Prompt processing depends on GPU cores, so Pro, Max and Ultra chips pull further ahead on long prompts.
Memory figures include an 8K-token KV cache and runtime overhead. Speeds assume memory-bound decoding and vary with the runtime, the quantisation format (MLX 4-bit vs GGUF Q4_K_M) and context length.