# LocalAIModelOnMac

> LocalAIModelOnMac is an independent, plain-English guide to running AI models locally on a Mac. It answers three questions: which Mac to buy (new, Apple Certified Refurbished, used or open-box), which open-weight model that Mac can run, and what model weights, parameter counts, mixture-of-experts "active" parameters and quantization (4-bit, 8-bit, GGUF, MLX) mean. Last updated Oct 2026. Not affiliated with Apple.

Key guidance:
- Memory (unified RAM) decides which models fit; memory bandwidth decides speed. RAM cannot be upgraded later.
- Rough sizes at 4-bit: about 0.55-0.6 GB per billion parameters, plus headroom for macOS and the conversation.
- 16GB Macs: models up to ~8-12B (e.g. Gemma 4 12B, Qwen3.5 9B, Liquid LFM2.5).
- 24-32GB: ~24-35B models (e.g. Qwen3.8 27B, Gemma 4 26B A4B; Gemma 4 31B and Qwen3.6 35B-A3B at 32GB).
- 48-64GB: the same models at 8-bit with long context.
- 96-128GB: ~100-125B MoE models (Qwen3.5 122B-A10B, Mistral Small 4, Llama 4 Scout).
- 192-256GB (Ultra Mac Studios only): very large MoE models (GLM-5.3-Flash, DeepSeek V4-Flash at 4-bit on 256GB, ~3-bit on 192GB). New M5 Ultra 256GB from $9,499; used M2 Ultra 192GB ~$3,700-5,950; used M3 Ultra 256GB ~$6,700-9,800 (Oct 2026).
- 512GB: GLM-5.3 tightly. Meta Muse Spark Open (816B total, 42B active) needs two M5 Ultra Mac Studios.
- Typical new prices (Apple US list, Oct 2026): Mac mini M6 16GB $899; Mac mini M5 Pro 64GB $2,699; Mac Studio M5 Max 128GB $5,099.
- Tools: LM Studio (app, GGUF + MLX), Ollama (app/CLI, local API), llama.cpp (GGUF engine), MLX / mlx-lm (Apple's framework).
- All figures are estimates; prices change, so check the current price with the seller.

## Pages
- [Home: which Mac runs which model](https://localaimodelonmac.com/): quick lookup table by memory tier
- [Mac buying guide](https://localaimodelonmac.com/macs/): memory tiers 16GB, 24-32GB, 48-64GB, 96-128GB, 192-256GB, 512GB with typical new and used costs, and where to find deals (Apple Certified Refurbished, Amazon, Best Buy open-box, eBay, Swappa)
- [Open-weight model directory](https://localaimodelonmac.com/models/): Qwen, Gemma, GLM, DeepSeek, Llama, Mistral, Muse Spark Open, Liquid LFM with size, active parameters, ~4-bit size and minimum Mac memory
- [Weights explained](https://localaimodelonmac.com/weights-explained/): parameters, MoE active parameters, quantization, GGUF vs MLX, memory math
- [Tools](https://localaimodelonmac.com/tools/): LM Studio, Ollama, llama.cpp, MLX/mlx-lm
- [About](https://localaimodelonmac.com/about/): methodology, disclosures, disclaimer