selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

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 →
#ModelParamsBest quantMemorySpeedLicence
136BA3BQ6_K≈31 GBFits in memory~61 tok/sCommercial OK
2
RecursiveSI logoflintRecursiveSI
39.8BA1.3BQ6_K≈33 GBFits in memory~146 tok/sCommercial OK
3
InternScience logoAgents-A1InternScience
35.1BA3.9BQ6_K≈29 GBFits in memory~47 tok/sCommercial OK
4
ByteDance-Seed logoSeed-OSS-36B-InstructByteDance-Seed
36.2BQ6_K≈32 GBFits in memory~5 tok/sCommercial OK
531.2BA4.6BQ6_K≈32 GBFits in memory~40 tok/sCommercial OK
6
XingChen-AGI logoXing4.0-29B-A4BXingChen-AGI
31.2BA4BQ5_K_M≈31 GBFits in memory~53 tok/sCommercial OK
733.1BQ6_K≈31 GBFits in memory~5.5 tok/sCommercial OK
8
InternScience logoSage-A1.5InternScience
36BA4.7BQ6_K≈30 GBFits in memory~38 tok/sCommercial OK
#ModelParamsBest quantMemorySpeedLicence
112.1BA2.5BQ8_0≈14 GBFits in memory~57 tok/sCommercial OK
232.8BQ6_K≈29 GBFits in memory~5.6 tok/sCommercial OK
334.7BA3.5BQ6_K≈31 GBFits in memory~53 tok/sCommercial OK
430.5BA3BQ8_0≈33 GBFits in memory~47 tok/sCommercial OK
530.5BA3.3BQ8_0≈33 GBFits in memory~43 tok/sCommercial OK
634BQ6_K≈29 GBFits in memory~5.4 tok/sConditions
739.8BQ5_K_M≈31 GBFits in memory~5.3 tok/sCheck licence
839.8BQ5_K_M≈31 GBFits in memory~5.3 tok/sCheck licence

Reasoning Models

All 70 that fit →
#ModelParamsBest quantMemorySpeedLicence
132.8BQ6_K≈29 GBFits in memory~5.6 tok/sCommercial OK
232.2BQ6_K≈28 GBFits in memory~5.7 tok/sCommercial OK
332.8BQ6_K≈29 GBFits in memory~5.6 tok/sCommercial OK
433.2BQ6_K≈29 GBFits in memory~5.5 tok/sCommercial OK
532.1BA3BQ6_K≈29 GBFits in memory~61 tok/sCommercial OK
633.2BQ6_K≈29 GBFits in memory~5.5 tok/sCommercial OK
714.8BQ8_0≈17 GBFits in memory~9.6 tok/sCommercial OK
832.1BA3BQ6_K≈29 GBFits in memory~61 tok/sCommercial OK

Vision-Language Models

All 413 that fit →
#ModelParamsBest quantMemorySpeedLicence
127.8BQ8_0≈31 GBFits in memory~5.1 tok/sCommercial OK
231.3BQ6_K≈33 GBFits in memory~5.8 tok/sCommercial OK
336BA3BQ6_K≈30 GBFits in memory~61 tok/sCommercial OK
427.8BQ8_0≈30 GBFits in memory~5.1 tok/sCommercial OK
525.8BA4BQ8_0≈29 GBFits in memory~35 tok/sCommercial OK
6
meta-models logoMuse-Glimmer-30Bmeta-models
29.8BQ8_0≈30 GBFits in memory~4.7 tok/sCommercial OK
736BA3BQ6_K≈29 GBFits in memory~61 tok/sCommercial OK
835.2BA3BQ6_K≈29 GBFits in memory~61 tok/sCommercial OK
#ModelParamsBest quantMemoryLicence
17.1BFP16 / BF16≈16 GBFits in memoryNon-commercial
27.1BFP16 / BF16≈16 GBFits in memoryNon-commercial
312.8BFP16 / BF16≈29 GBFits in memoryConditions
420.4BQ8_0≈25 GBFits in memoryCommercial OK
5
Tongyi-MAI logoZ-Image-TurboTongyi-MAI
6.2BFP16 / BF16≈14 GBFits in memoryCommercial OK
6
OpenMOSS-Team logoMOSS-Image-SFTOpenMOSS-Team
5.1BFP16 / BF16≈12 GBFits in memoryCommercial OK
712.8BFP16 / BF16≈29 GBFits in memoryConditions
820.4BQ8_0≈25 GBFits in memoryCommercial OK

Speech-to-Text

All 100 that fit →
#ModelParamsBest quantMemoryLicence
12.1BFP16 / BF16≈5.1 GBFits in memoryCommercial OK
22.3BFP16 / BF16≈5.8 GBFits in memoryCommercial OK
3627MFP16 / BF16≈1.9 GBFits in memoryCommercial OK
48.7BFP16 / BF16≈20 GBFits in memoryCommercial OK
5
microsoft logoVibeVoice-ASRmicrosoft
8.7BFP16 / BF16≈20 GBFits in memoryCommercial OK
6473MFP16 / BF16≈1.6 GBFits in memoryCommercial OK
72.3BFP16 / BF16≈5.7 GBFits in memoryCommercial OK
84.4BFP16 / BF16≈10 GBFits in memoryCommercial OK

Text-to-Speech

All 113 that fit →
#ModelParamsBest quantMemoryLicence
1
m-a-p logoYuE2-3Bm-a-p
3.6BFP16 / BF16≈8.6 GBFits in memoryNon-commercial
2
BreezeBlue logoBreeze-TTS-2BreezeBlue
3.5BFP16 / BF16≈8.3 GBFits in memoryNon-commercial
3
openbmb logoVoxCPM2openbmb
2.3BFP16 / BF16≈5.6 GBFits in memoryCommercial OK
4
k2-fsa logoOmniVoicek2-fsa
613MFP16 / BF16≈1.9 GBFits in memoryCheck licence
51.9BFP16 / BF16≈4.8 GBFits in memoryCommercial OK
61.5BFP16 / BF16≈3.8 GBFits in memoryCommercial OK
7915MFP16 / BF16≈2.5 GBFits in memoryCommercial OK
8
fishaudio logos2-profishaudio
4.6BFP16 / BF16≈11 GBFits in memoryNon-commercial

Embedding Models

All 244 that fit →
#ModelParamsBest quantMemoryLicence
1744MFP16 / BF16≈2.2 GBFits in memoryCommercial OK
2303MFP16 / BF16≈1.2 GBFits in memoryConditions
37.6BFP16 / BF16≈17 GBFits in memoryCommercial OK
4596MFP16 / BF16≈1.8 GBFits in memoryCommercial OK
54BFP16 / BF16≈9.5 GBFits in memoryCommercial OK
68BFP16 / BF16≈18 GBFits in memoryCommercial OK
71.1BFP16 / BF16≈3.0 GBFits in memoryCommercial OK
8
BAAI logobge-m3BAAI
558MFP16 / BF16≈1.8 GBFits in memoryCommercial OK

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.

MLX Apple Silicon with mlx-lmdocs
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 OpenAI-compatible serverdocs
mlx_lm.server --model mlx-community/Ornith-1.5-35B-A3B-OptiQ-4bit --port 8080

API on http://localhost:8080/v1.

Ollama From the Ollama librarydocs
ollama run ornith-1.5:35b

The library build picks its own default quantisation.

Ollama Q6_K straight from Hugging Facedocs
ollama run hf.co/bartowski/Ornith-1.5-35B-A3B-GGUF:Q6_K
llama.cpp OpenAI-compatible server with a web UIdocs
llama-server -hf bartowski/Ornith-1.5-35B-A3B-GGUF:Q6_K

brew install llama.cpp ships with Metal. Downloads the GGUF on first run.

LM Studio Download with the lms CLIdocs
lms get bartowski/Ornith-1.5-35B-A3B-GGUF@Q6_K

Then 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.