Running Local Open-Weight LLMs on a Mac: Which Model Fits Which Mac, and How Fast (Aug 2026)

Phi-4, Gemma 3, Qwen3, Llama 4, DeepSeek V4, Kimi K2 — 8 mainstream open models × Apple Silicon, computing the minimum Mac + memory + estimated speed for inference. Key: MoE models make Macs punch above their weight, but 1T-class flagships exceed Mac memory limits even at Q4 — they simply won't run locally.

ModelType (total / active)Q4 sizeMin Mac + memoryEst. speed (ceiling)
Phi-4 14Bdense 14B8.9 GBM5 base 32GB~17 tok/s (real ~10–14)
Gemma 3 27Bdense 27B16.5 GBM5 Pro 64GB~19 tok/s (real ~11–15)
Qwen3.6 27Bdense 27B16.8 GB (2nd-hand)M5 Pro 64GB~18 tok/s (real ~11–15)
Llama 4 ScoutMoE 109B / 17B active~63 GB (est)M5 Max 128GB~65 tok/s (17B active; real ~40–52)
Qwen3 235B-A22BMoE 235B / 22B active142 GB (measured)M5 Ultra 256GB~100 tok/s (22B active; real ~60–80)
GLM-4.6MoE 355B / 32B active~200 GB (est)M5 Ultra 256GB~67 tok/s (32B active; real ~40–54)
Kimi K2MoE 1T / 32B active~570 GB (est)⚠ exceeds 512GB ceilingWon't run locally (over budget even at Q4)
DeepSeek V4-ProMoE 1.6T / 49B active~800 GB (est)⚠ exceeds 512GB ceilingWon't run locally (use V4-Flash 284B/13B variant)

Method & sources

**Nature: this is an engineering estimate, not a measured benchmark.** Two verified inputs: (1) model params/quant sizes from each HuggingFace model card (Q4 sizes for Qwen3-235B, Gemma 3, Phi-4 are measured GGUF; those for DeepSeek V4-Pro, Kimi K2, Llama 4 Scout, GLM-4.6 are estimates, varying by tens of GB depending on quant method); (2) per-chip memory ceiling and bandwidth from Apple (M5 153 / M5 Pro 307 / M5 Max 614 / M5 Ultra 1200 GB/s; M6 is a new base at 170GB/s but Mac-mini-only, 32GB ceiling). **Formulas (verifiable yourself)**: (i) usable memory ≈ unified memory ×0.75; (ii) model footprint ≈ Q4 file size + KV cache (grows with context); (iii) **est. speed (tok/s) ≈ memory bandwidth ÷ Q4 size of ACTIVE params** — MoE counts only active params (why Macs run MoE fast). **⚠ Speed is a theoretical ceiling: real throughput is affected by prompt length, KV cache, quant method and software (MLX/llama.cpp), and typically lands at 60–80% of the estimate, lower for long context.** License note: latest Kimi K2.7 / GLM-4.6 licenses & params weren't verified verbatim; base-version figures used — confirm at the official repo before relying on them. Training/fine-tuning is out of scope (Macs suit inference and light QLoRA; use NVIDIA GPUs for serious training). Sources: Apple https://www.apple.com/mac-studio/specs/ ; Qwen3-235B https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507 ; DeepSeek V4-Pro https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro ; Kimi K2 https://huggingface.co/moonshotai/Kimi-K2-Instruct .

Source: https://www.apple.com/mac-studio/specs/

Retrieved: 2026-08-26

FAQ

What does "Running Local Open-Weight LLMs on a Mac: Which Model Fits Which Mac, and How Fast (Aug 2026)" cover?
It covers Phi-4 14B, Gemma 3 27B, Qwen3.6 27B, Llama 4 Scout, Qwen3 235B-A22B, GLM-4.6, Kimi K2, DeepSeek V4-Pro, compared across: Model, Type (total / active), Q4 size, Min Mac + memory, Est. speed (ceiling).
What are the sources and methodology?
**Nature: this is an engineering estimate, not a measured benchmark.
When was this data last updated?
The data was retrieved/updated on 2026-08-26.

Related datasets