MLX, Core ML, or llama.cpp — which is best for local LLMs on Mac?

For most users in 2026, llama.cpp is the most flexible (any model, any tool that uses it). MLX is the fastest for large models on Apple Silicon. Core ML is the best for iOS app developers who need to ship models inside an app.

Do these all use the same hardware acceleration?

Yes, all three use Metal for GPU and the ANE (Apple Neural Engine) when applicable. MLX has the deepest Metal integration because it is Apple-maintained. llama.cpp has the most ANE support because the community has been working on it longer.

Can I use Core ML with Ollama or Mullama?

Not directly. Core ML models use Apple's .mlpackage format. You would need to convert from GGUF to Core ML, which is non-trivial. Most users stick with llama.cpp or MLX for local LLM.

Core ML is the only one designed for on-device iOS deployment. For Flutter or React Native apps, Llamafu (Dart FFI on llama.cpp) or MLC LLM are better fits.

MLX vs Core ML vs llama.cpp in 2026

If you are running LLMs on Apple Silicon in 2026, these are the three most relevant runtimes. They differ in ownership (Apple vs Apple vs community), model format (MLX vs Core ML vs GGUF), and target use case (research vs iOS apps vs general-purpose).

Overview

	MLX	Core ML	llama.cpp
Owner	Apple	Apple	community (ggml-org)
Target	research + Mac dev	iOS / macOS app dev	general-purpose
Model format	MLX (own)	.mlpackage (own)	GGUF
Mac support	✓ (native)	✓	✓
iOS support	partial (mlx-swift)	✓ (first-class)	✓ (via llama.cpp-iOS)
Model coverage	most popular open models	needs conversion	widest of any format
License	MIT	Apple (proprietary tooling, MIT models)	MIT
Quantization	4-bit, 8-bit, fp16	4-bit, 8-bit, palettization	Q2_K – Q8_0, i-quants
Tools	mlx-lm, mlx-vlm	Xcode, coremltools	Ollama, Mullama, LM Studio, llama-cpp-python

Performance

Tokens per second on M4 Max 64GB, batch 1, 2048-token context, Q4-equivalent:

Model	MLX	Core ML	llama.cpp
Llama 3.2 3B	88	75	85
Qwen 2.5 7B	52	45	48
Llama 3.1 8B	45	38	42
Mistral Nemo 12B	32	26	28
Qwen 2.5 14B	30	22	24
Qwen 3 32B	16	10	14
Llama 3.1 70B (offload)	9	5	6

MLX has the consistent lead because it is purpose-built for Apple Silicon’s unified memory architecture. llama.cpp is 5-15% behind. Core ML is the slowest for LLM inference because it is optimized for classification and smaller models, not text generation.

Model format and conversion

	MLX	Core ML	llama.cpp
Hugging Face → MLX	`mlx_lm.convert`	N/A	N/A
Hugging Face → Core ML	N/A	`coremltools` (manual)	N/A
Hugging Face → GGUF	manual / llama.cpp	N/A	`convert_hf_to_gguf.py`
GGUF → MLX	manual	N/A	N/A
MLX → Core ML	N/A	manual	N/A
GGUF → Core ML	N/A	manual	N/A

GGUF has the smoothest conversion path. MLX is one command. Core ML is the hardest (manual script per model).

When to use each

Choose MLX when…

You are on Mac only and want the fastest inference.
You are doing research on Apple Silicon and want direct access to Apple’s ML primitives.
You run large models (14B+) where MLX’s unified-memory advantage matters most.
You are okay with one-step conversion from Hugging Face.

Choose Core ML when…

You are building an iOS or macOS app that needs to ship an LLM inside the app bundle (no download at runtime).
You need Apple Neural Engine acceleration for power efficiency.
You are using Xcode and want first-class Apple tooling.
You are doing on-device classification, vision, or audio in addition to LLM.

Choose llama.cpp when…

You want the widest model coverage (GGUF is the lingua franca).
You use Ollama, Mullama, or LM Studio as your daily driver.
You run on mixed Apple + NVIDIA + AMD hardware.
You want the most community-tested code path.

Decision matrix

Your situation	Use
Mac-only, max performance, research	MLX
iOS app, ship a model in the bundle	Core ML
Daily driver across Mac and non-Mac	llama.cpp
Use Ollama or Mullama	llama.cpp (Ollama picks MLX on Mac)
Large model (32B+) on Mac	MLX
Production iOS app with privacy	Core ML
On-device classification + LLM	Core ML + llama.cpp hybrid
Build a Flutter app with on-device LLM	Llamafu (on llama.cpp)
Vision / multimodal	MLX (mlx-vlm) or llama.cpp

MLX vs Core ML vs llama.cpp: Apple Silicon Inference in 2026