Local LLMs
Best local LLM for a 16 GB Mac
On 16 GB, plan for about 8–9 GB of usable memory once macOS and your apps are counted, which means an 8B model at 4-bit is the sweet spot. A 4B model is faster and lighter; a 14B model technically fits but leaves no room for context and starts swapping. Pick Qwen3 8B for general work, DeepSeek R1 Distill 8B for reasoning.
16 GB is the most common Mac configuration sold — and the most misunderstood. It is enough for a genuinely useful local model, but not enough to run whatever looks impressive on a leaderboard.
Here is the honest math.
Start by subtracting, not adding
16 GB is not a 16 GB budget for the model. Roughly:
- 4–6 GB goes to macOS, the window server and background services.
- 2–3 GB goes to whatever you actually have open — a browser is not free.
- 1–3 GB should stay free so the system does not swap.
That leaves around 8–9 GB for weights plus KV cache. Anything above that starts borrowing from the swap file, and swap on an SSD is not slow so much as brutal for token generation.
What fits
| Model class | 4-bit size | Verdict on 16 GB |
|---|---|---|
| 4B (Qwen3 4B, Gemma 3 4B) | ~2.5–3 GB | Comfortable, fast, good for everyday chat |
| 8B (Qwen3 8B, Llama 3.2 8B) | ~5 GB | The sweet spot — good quality, responsive |
| 12B (Gemma 3 12B) | ~8 GB | Works, but leaves little room for long context |
| 14B (Qwen3 14B, Phi-4) | ~9 GB | Only with a short context and nothing else open |
| 24B+ | 14 GB+ | Do not. It will swap. |
Why 8B beats 14B on this machine
A 14B model at 4 bits is not twice as good as an 8B model — and on 16 GB it is often worse in practice, for two reasons:
- Context gets squeezed. With 9 GB of weights resident, a 32k context can push you into swap. You end up dropping to 4k context, which hurts real work more than the extra parameters help.
- Speed falls off a cliff once it swaps. A model generating 2 tokens per second is unusable regardless of its benchmark score.
The reliable pattern on 16 GB: pick the 8B class, give it a long context, and keep it loaded.
Concrete picks
- Everyday chat and writing — Qwen3 8B. Strong instruction following, small footprint, fast on M-series.
- Coding — Qwen3 8B or a code-tuned 7B. Below 14B, coding quality depends more on how you prompt than on the extra parameters.
- Reasoning — DeepSeek R1 Distill 8B. It shows its working, which makes it useful for maths even when it is wrong; on 16 GB do not go to the 14B distill.
- Vision — Qwen2.5-VL 7B or Gemma 3 4B. Small multimodal models are genuinely usable for screenshots and document Q&A.
- Image generation — this is a different budget. SDXL in ComfyUI will run on 16 GB but is slow; see ComfyUI on Apple Silicon.
The three settings that matter more than the model
- Keep the model loaded.
OLLAMA_KEEP_ALIVE=30m— reloading a 5 GB model every time is the biggest waste of your day. - Cap the context. 8k is plenty for chat and code review; 32k is what pushes you into swap.
- Close the browser during long runs. Unglamorous, effective.
When 16 GB is genuinely not enough
If your work involves long documents (50+ pages), multi-file codebases, or iterating on image models, 16 GB will feel like a wall — not because the models are bad, but because context is the first thing you run out of. That is a hardware conversation, not a software one.
Check your exact machine with the model checker, and if you are new to this, start with Ollama on Mac.
FAQ
Can a 16 GB Mac run a 14B model?
It can load one, but it will not feel good. A 14B model at 4 bits is around 9 GB, which is most of your usable memory before any context is added. Short prompts may work; a long conversation will push macOS into swap and generation becomes painfully slow.
Is 16 GB enough for coding with a local model?
For autocomplete, small refactors and explaining code, yes — an 8B model is genuinely useful. For whole-repository reasoning, the limit is context length rather than raw intelligence, and 16 GB runs out of context before it runs out of model quality.
Does more RAM or a faster chip matter more?
RAM decides what you can run; bandwidth decides how fast it feels. On 16 GB machines, a faster chip makes the same 8B model snappier but does not let you run a bigger one.
Will Apple’s own on-device models make this unnecessary?
Partially. System-level models will handle small, private tasks well. They will not replace a 14B model you chose yourself for a specific job, and they will not run third-party weights — so the memory maths on this page stays relevant.