I have been trying local language models on a MacBook Air with 24GB of unified memory.
For token generation, memory bandwidth matters more than total RAM. That is why Apple computers can be faster than other unified-memory platforms: the 13-inch MacBook Air M4 has 120GB/s of memory bandwidth, while the 16-inch MacBook Pro with M5 Pro—around $2,800 in the US, though Apple’s pricing may change—has about 307GB/s. AMD’s Strix Halo platform can provide up to 128GB of unified system memory, but its 256GB/s bandwidth is the more relevant number when comparing token-generation speed.
With memory bandwidth in mind, these are the models I recommend for this hardware, in no particular order.
unsloth/gemma-4-12B-it-qat-GGUFunsloth/gemma-4-26B-A4B-it-GGUFunsloth/gemma-4-E2B-it-GGUFunsloth/gemma-4-E4B-itunsloth/Qwen3.5-4B-MTP-GGUFunsloth/Qwen3.5-9B-MTP-GGUF
The Gemma models are better at prose, while the Qwen models are for agentic tasks.
I started with Unsloth because it automatically configured models and tooling particularly well. Tool calling was smooth and never fell into infinite tool-call loops, unlike LM Studio at the time of my May 2026 testing. I did run into an issue with Unsloth that raised possible security concerns about its Python sandbox, though.
I eventually moved to llama.cpp because it gives me a smaller, more focused runtime with fewer moving parts. Since November 2025, it has included a SvelteKit-based WebUI; see the discussion.
For a 24GB MacBook Air M4, I still recommend trying Unsloth’s GGUF models. Their quantizations can offer a strong quality-to-memory tradeoff, and smaller files may generate tokens faster than larger official releases. I would not treat Unsloth as automatically better, though: compare the exact model, quantization, and runtime when an official release is available.
Since Qwen3.8 came out, I’ve been experimenting with llama.cpp’s native model-downloading functionality. That’s the setup I’m a fan of for my 48GB MacBook Pro M5.