Home › Mac › Mac mini (M4, 2024) 24GB
Best Local LLM for a Mac mini M4 (24GB, 2026): What Fits
The first paid memory step Apple lists for this chip. In our own dataset the budget gain is concrete: the usable-memory rule gives 16 GiB here against 10.7 GiB on the 16GB machine. Apple lists this machine’s memory as “Apple M4 chip · 16GB unified memory · Configurable to: 24GB or 32GB” (spec-sheet line breaks shown as ·)[1]. “Unified” is not the same as “available”: MLX’s docs describe the architecture plainly — “Apple silicon has a unified memory architecture. The CPU and GPU have direct access to the same memory pool”[3] — the same pool the operating system lives in. The macOS GPU-usable fraction Apple does not publish (no public data (we checked, 2026-10-23)), so the tables below run on our stated heuristic: 24GB → 16.0 usable GiB (rule and reasoning).
Spec check date: 2026-10-23 — memory configurations are quoted from Apple’s own tech-spec pages fetched that day; runtime statements come from each project’s docs fetched the same day (links at the bottom). Where a vendor is silent the page says so instead of guessing.
What fits (sourced weights + KV cache, 8k context)
| Model | Best sensible quant | Needs (GiB) | Budget (GiB) |
|---|---|---|---|
| Gemma 4 31B | q2_k | 15.9 | 16.0 |
| gpt-oss-20b | q5_k_m | 15.9 | 16.0 |
| Muse Glimmer 30B | q3_k_m | 15.6 | 16.0 |
| Qwen3.6 27B | q3_k_m | 15.0 | 16.0 |
| Qwen3.8 27B | q3_k_m | 15.0 | 16.0 |
| Qwen3 30B-A3B (MoE) | q2_k | 14.4 | 16.0 |
Sizes: modelvram.com open dataset (CC BY 4.0 / MIT, retrieved 2026-10-20), KV extrapolated by the published formula. 20 dataset models fit under the 16.0 GiB budget.
The near misses, printed as misses
| Model | Lowest sourced quant | Needs (GiB) | Verdict |
|---|---|---|---|
| Ornith 1.5 35B-A3B (MoE) | q2_k | 16.1 | over by ≈0.1 GiB |
| Qwen3.6 35B-A3B (MoE) | q2_k | 16.1 | over by ≈0.1 GiB |
| Llama 3.1 70B | q2_k | 33.5 | over by ≈17.5 GiB |
Quantisation shrinks GiB, not a model’s category: if the model you want is three tiers above budget, the honest answer for a 24GB Mac is “not this machine” — quant tables per model or the calculator.
Which runtime covers Apple silicon
- Ollama — “Ollama supports GPU acceleration on Apple devices via the Metal API” — a section of its own in the GPU docs; no per-chip memory guidance there[6]
- llama.cpp — README: “Apple silicon is a first-class citizen” (ARM NEON, Accelerate, Metal); Metal listed for Apple Silicon in its backend table; build docs: “On MacOS, Metal is enabled by default”[5]
- LM Studio — Docs: on Apple Silicon Macs it can run llama.cpp models or “Apple’s MLX”; its system-requirements line reads “LM Studio generally supports Apple Silicon Macs, x64/ARM64 Windows PCs, and x64 Linux PCs”[7]
- vLLM — Listed as a GPU variant, but the install page routes it to vLLM-Metal — “a community-maintained hardware plugin that uses MLX as the compute backend”, not core vLLM[8]
Side-by-side, cell by cell, each claim linked to its docs line — all four engines in one table: runtime comparison table · Ollama vs llama.cpp · Ollama vs LM Studio · Ollama vs vLLM · vLLM vs llama.cpp.
Making the most of 24GB
- Context is the hidden cost. Every 1k tokens adds KV cache on top of weights; the table is 8k context and the calculator scales the KV term if you plan longer.
- The OS is a passenger you cannot evict. Apps and a browser draw from the same pool — which is why this page budgets 33% for system + headroom instead of selling you the full 24GB.
- If it does not fit locally, that is a cost question. under $500 · under $1500 · cloud GPU compares.
FAQ
How much of the unified memory can a model actually use?
Apple does not publish a GPU-usable figure for any Mac — no public data (we checked, 2026-10-23). Its Metal documentation names the mechanism (“keep the total memory footprint of its resources and heaps less than this threshold”) without giving the number, so this page uses our own rule: two thirds of unified memory up to 36GB, three quarters above it — here 24GB → 16.0 usable GiB. A heuristic, not a vendor spec; check your machine with sysctl iogpu.wired_limit_mb.
Does it run on Ollama, LM Studio, llama.cpp or vLLM?
All four document Apple silicon one way or another — Ollama via the Metal API, llama.cpp with Metal on by default, LM Studio choosing between llama.cpp (GGUF) and MLX, vLLM only through the community vLLM-Metal plugin. Each claim is quoted with its docs link in the runtime section above (fetched 2026-10-23).
What is the biggest model that fits a 24GB Mac?
From our dataset (46 models, weights + KV at 8k context): 20 models fit under the usable-memory rule above. The table above is sorted by how much they need; the tightest one that fits is the honest ceiling for a 24GB Mac at 8k context — raise the context length and the KV-cache term grows with it.
Why does this page have no tokens/s numbers?
Because none of the four runtimes publishes an Apple-silicon throughput table we could verify, and third-party Mac benchmarks mix chips, quants and context into numbers nobody can reproduce. Fit is decidable arithmetic; speed on your specific machine is a measurement you have to take. We print the arithmetic and leave the rest blank.
Sources for this page
- Apple Support — Mac mini (2024) Tech Specs — fetched 2026-10-23
- MLX docs — Unified Memory — fetched 2026-10-23
- Apple Developer — MTLDevice.recommendedMaxWorkingSetSize — fetched 2026-10-23
- llama.cpp docs — build (Metal Build section) — fetched 2026-10-23
- Ollama docs — GPU support — fetched 2026-10-23
- LM Studio docs — fetched 2026-10-23
- vLLM docs — GPU installation — fetched 2026-10-23
Other Mac tiers we have pages for: Mac mini (M4, 2024) 16GB · Mac mini (M4 Pro, 2024) 48GB · MacBook Pro (14-inch, M4, 2024) 16GB. Not listed here? The calculator covers 46 models at any context length.