Local Models
Local language, vision and audio models tested on Apple Silicon: Qwen3, Gemma3, Llama, Mistral and more — benchmarks and RAM requirements for M1–M4 Macs.
- Find the right model
- Setup per model
- Benchmark comparisons
- Know RAM needs
What counts as a local model?
Runs on your Mac
The model weights are downloaded and inference runs locally through Ollama, LM Studio, MLX, llama.cpp or a similar runtime.
Open weights, not always open source
Many local models are open-weight, but their license may still restrict commercial use, redistribution or fine-tuning.
Memory matters
Model size is not total memory use. Context length, KV cache, quantization, vision input and other apps also affect unified memory.
Privacy depends on configuration
Local inference can keep prompts on your Mac, but downloads, plugins, cloud features, exposed local servers and backups can still create data paths.
Start here
Local model checklist
- Is the model actually downloadable?
- Does it have Ollama, GGUF, MLX or LM Studio support?
- Is it text-only, vision-capable, audio-capable, or multimodal?
- What license applies: open source, open weights, research-only or commercial?
- How much unified memory is realistic after context and KV cache?
- Does it need cloud features, API calls or online tools?
- Can you run it offline after download?
- Does it fit your task better than a smaller model?
-
Qwen3.6-27B Fable Fusion 711 on Mac: GGUF Sizes and RAM Choices
Fable Fusion 711 on Mac: start with the 18.0 GB Q4_K_M GGUF on 32 GB. File sizes for 24 to 64 GB, MTP and vision files, no original Mac benchmarks.
-
Apple Intelligence Local AI: On-Device Models, PCC and Apple Silicon
Which Apple Intelligence requests stay on your Mac, which go to Private Cloud Compute or ChatGPT, and which need an M3 with 12GB.
-
Gemma 4 26B vs Qwen3.6 27B on Mac: RAM, Ollama and MLX
Gemma 4 26B A4B and Qwen3.6 27B compared on Mac: RAM, Ollama, MLX, coding, multimodality and the newer Qwen3.8 successor.
-
Best Open-Weight LLMs for Mac in 2026: RAM, Ollama Tags & Picks
Open-weight LLM picks for 8–64GB Macs: Gemma 4 for smaller systems, Qwen3.8-27B for 32GB+, Ollama tags, realistic memory margins and a fair speed test.
How local model recommendations are made
Local model recommendations on AI on Mac should separate model size, quantization, runtime, context length, Apple Silicon generation and unified memory. A model that works on a 48 GB Mac Studio may be unrealistic on an 8 GB MacBook Air. The articles in this category should also distinguish between open source, open weights, cloud-only APIs and hybrid tools.