M4 Mac Mini restocks; buyers debate whether 24GB RAM is enough for local AI

Apple’s Mac mini lineup is feeding a new buyer argument that has less to do with desktop computing than with local AI. As M4 Mac mini availability appeared to improve in user posts, the practical question shifted to memory: whether 24GB is enough for running local large language models, or whether buyers should pay up for higher unified-memory tiers on M4 Pro machines. Apple’s current Mac mini pages show the standard M4 model configurable from 16GB to 24GB unified memory, while M4 Pro configurations can go to 48GB. The argument is rooted in how Apple Silicon handles AI workloads. Apple uses unified memory, which means the CPU and GPU draw from the same pool rather than from separate system RAM and VRAM. On the current Mac mini, Apple lists 120GB/s memory bandwidth for M4 and 273GB/s for M4 Pro, a difference that matters for users trying to keep larger quantized models in memory and sustain inference speed. ### Why are people fixated on 24GB? (apple.com) Apple’s $999 Mac mini configuration pairs the M4 chip with 24GB of memory and 512GB of storage, putting that setup near the center of the discussion for buyers who want a relatively low-cost local AI box. Apple’s store also shows the base M4 model starting at $799 with 16GB, with 24GB as the top memory option unless the buyer moves up to a different chip. (apple.com) Google’s Gemma 4 documentation helps explain why 24GB has become a dividing line. Google says Gemma 4 models are available in multiple sizes, including E2B, E4B, 26B A4B and 31B, and that lower-precision quantization can reduce memory requirements. The company also says higher-parameter and higher-precision models are generally more capable but cost more in memory and compute. ### Can 24GB actually run local models? (apple.com) Google says quantized Gemma models can run with 16-bit, 8-bit or 4-bit representations, and specifically notes that frameworks such as llama.cpp and Ollama make it possible to run versions of Gemma on a laptop or other small device without a discrete GPU. That means a 24GB Mac mini can run some local models, especially smaller or aggressively quantized ones. Google’s own guidance points developers toward the Gemma 4 26B A4B model as “a good place to start” because it has lower resource requirements, while still targeting many general tasks. (ai.google.dev) The company also says the model family spans deployments from mobile and edge devices to consumer GPUs and workstations, underscoring that not every Gemma 4 variant is aimed at the same class of hardware. ### So why do some buyers want 64GB or 128GB? (ai.google.dev) Apple’s current Mac mini lineup does not offer 64GB or 128GB memory on the configurations visible in its Mac mini technical specifications and store pages reviewed here. The standard M4 tops out at 24GB, and the M4 Pro tops out at 48GB. Buyers talking about 64GB to 128GB for local AI are usually describing a broader requirement for heavier inference workloads, larger models, more headroom for context windows, or running multiple services at once, rather than a currently listed Mac mini option. (ai.google.dev) Google’s model overview supports that distinction. The company says larger Gemma 4 models target “consumer GPUs and workstations,” and says memory needs vary by model size, precision and environment. In practice, that means the answer to “is 24GB enough?” depends on whether the user wants a responsive local assistant, a coding model, multimodal experiments, or sustained production-style inference. ### What is the cleanest takeaway for buyers? (apple.com) Apple’s current product pages support a narrow conclusion: 24GB is the ceiling for an M4 Mac mini, and 48GB is the ceiling for an M4 Pro Mac mini. Google’s current Gemma 4 documentation supports a second one: quantization broadens what can run locally, but capability and speed rise with model size, precision and memory. Apple’s Mac mini comparison and configuration pages, along with Google’s Gemma 4 model overview and run guides, are the most direct references for the next step: matching a named Mac mini memory tier to a specific local model and quantization target before buying. (ai.google.dev) (apple.com 1) (apple.com 2)