Memory fit check
Can my Mac run a local LLM?
A direct answer to the most common question before a local AI Mac purchase: will this model actually fit this machine? The table uses the same capacity arithmetic as our fit explorer — public weight sizes against usable unified memory — so you can see the lowest Mac tier for each current model.
Short answer
A Mac can run a local model only if the stored weights fit inside usable unified memory after macOS, the runtime, and context reserves. Planning rule: about 75% of unified memory is usable, keep 4GB for runtime and context, and a 4-bit quantization needs roughly 0.6GB per billion parameters. Beyond that arithmetic you are choosing speed and workflow, not capacity.
The lowest Mac tier for each current model
Every row reuses the fit engine capacity arithmetic: usable memory is 75% of unified memory minus a 4GB runtime and context reserve, and a tier is marked tight when less than 4GB remains. The first tier with real headroom uses the same threshold as the buying-decision card in the fit explorer. Tap any model name to open it in the interactive explorer.
| Model | Weights (quantization) | Lowest tier that fits | First tier with real headroom | 16GB Mac mini M6 |
|---|---|---|---|---|
| Qwen3.8 27B (27B) | 16.06GB · verified file size (Q4_0) | Mac mini M6 · 32GB · Tight fit · 3.94GB GB left | Mac mini M5 Pro · 48GB · 15.94GB GB left | Does not fit |
| Qwen3.8-Flash-Next 180B (6B active) (180B total · 6B active) | 360GB · verified file size (BF16) | Mac Studio M5 Ultra · 512GB · Headroom · 20GB GB left | Mac Studio M5 Ultra · 512GB · 20GB GB left | Does not fit |
| Qwen3.8 2.4T-A95B (2400B total · 95B active) | 1440GB · planning estimate (Q4_K_M) | No catalog Mac fits | — | Does not fit |
| DeepSeek-V4-Flash 284B-A13B (284B total · 13B active) | 170.4GB · planning estimate (MLX-4bit) | Mac Studio M5 Ultra · 256GB · Headroom · 17.6GB GB left | Mac Studio M5 Ultra · 256GB · 17.6GB GB left | Does not fit |
| DeepSeek-V4-Pro 1.6T-A49B (1600B total · 49B active) | 960GB · planning estimate (MLX-4bit) | No catalog Mac fits | — | Does not fit |
| GLM-5.3-Flash 320B-A18B (320B total · 18B active) | 203.99GB · verified file size (MLX-4bit) | Mac Studio M5 Ultra · 512GB · Headroom · 176.01GB GB left | Mac Studio M5 Ultra · 512GB · 176.01GB GB left | Does not fit |
| GLM-5.3 744B-A40B (744B total · 40B active) | 446.4GB · planning estimate (MLX-4bit) | No catalog Mac fits | — | Does not fit |
| Gemma 4 E2B (5.1B) | 3.35GB · verified file size (Q4_0) | Mac mini M6 · 16GB · Headroom · 4.65GB GB left | Mac mini M6 · 16GB · 4.65GB GB left | Headroom |
| Gemma 4 12B Unified (11.95B) | 6.98GB · verified file size (Q4_0) | Mac mini M6 · 16GB · Tight fit · 1.02GB GB left | Mac mini M6 · 24GB · 7.02GB GB left | Tight fit |
| Gemma 4 26B-A4B (25.2B total · 3.8B active) | 14.44GB · verified file size (Q4_0) | Mac mini M6 · 32GB · Headroom · 5.56GB GB left | Mac mini M6 · 32GB · 5.56GB GB left | Does not fit |
How the fit arithmetic works
- Stored weights gate capacity. The model file must fit in memory before anything else matters. For MoE models, the full stored package counts — active parameters do not shrink the requirement.
- Usable memory is less than the sticker number. Budget 75% of unified memory for the model, then keep 4GB for macOS, the runtime, and a working context. A 16GB Mac therefore plans around 8GB for weights.
- Tight versus headroom. If less than 4GB remains after the weights, long prompts, agents, and concurrent jobs can push the setup over the line. That tier is labeled tight, not headroom.
When the answer is no
If the model does not fit, a faster chip will not change the answer — memory is the gate. Your realistic options, in order of cost: pick a smaller quantization of the same model, step down to a smaller or distilled model that still does the job, or keep a cloud API for that workload and reserve the Mac for models that fit. Do not buy a maximum-memory Mac Studio for one model whose Mac runtime is still unverified; check the runtime evidence first.
Interactive tool
Check your exact configuration.
Pick your model, quantization, and runtime to see the weight budget, remaining memory, and the next step into the free buy / keep / wait advisor.
Continue the decision
- Best Mac for local LLMsThe full buying guide for memory, context, runtime, and sustained work.
- Local model evidence matrixPer-model artifact sizes, runtime evidence, and capability labels.
- Mac mini M6 specsPrice, release date, memory tiers, and the entry-tier local AI limits.
- Mac mini vs Mac StudioWhich constraints each machine actually removes.
Questions people ask
Clear answers before you buy
Can a 16GB Mac run a 70B model?
No. A 70B model at 4-bit quantization stores roughly 42GB of weights, while a 16GB Mac has a planning budget of about 8GB after the system, runtime, and context reserves. This is not a speed problem; the weights do not fit. Even a 32B model at Q4 (about 20GB of weights) is tight on a 32GB Mac after reserves.
Do MoE active parameters reduce the memory I need?
No. Mixture-of-experts models still store every expert in memory. Active parameters explain per-token compute, not capacity. The gate is the full stored package, which is why a 284B-A13B model needs far more memory than its 13B active parameters suggest.
Is "fits with headroom" a promise it will run well?
No. Fit only answers whether the weights fit the memory budget. Prompt processing speed, long context, tool calling, concurrency, and sustained behavior depend on the exact model, runtime (MLX, Ollama, llama.cpp), and client. Treat runtime evidence as a separate question.
Which Mac should I buy for local LLMs in 2026?
Start from your model, not the chip name. Use the table above to find the minimum tier with real headroom, then run the free advisor with your context length, concurrency, and budget. The full decision guide compares Mac mini M6, Mac mini M5 Pro, and Mac Studio by workload.
Are the weight sizes verified?
Rows marked "verified file size" use exact public artifact bytes from official or community repositories, checked on 2026-08-28. Rows marked "planning estimate" use the labeled 0.6GB-per-billion-parameter rule and stay estimates until an exact file is read back.
Use your own baseline
Get a free recommendation for your actual bottleneck.
The advisor asks for the hardware, workload, budget, and compatibility details that generic buying guides cannot see.
Keep or Upgrade is an independent decision-support tool. Apple product names identify the products being compared; Apple does not sponsor or endorse this site. Specifications and vendor test claims can change, so confirm local availability, pricing, and compatibility before purchasing.