WIRED Review Finds Memory Limits Apple’s M6 Mac Mini for AI Agents

Apple’s M6 Mac Mini arrives with an AI-heavy pitch, including new Neural Accelerators in every GPU core and Apple’s claim of twice the peak compute for on-device AI workflows. But WIRED’s hands-on review finds a more constrained reality: the $899 base machine is responsive with smaller local models and can generate images locally, while its 16 GB of memory becomes a hard limit for larger models and a basic AI-agent workflow.

That distinction matters because local AI is not one workload. A model answering conversational prompts has different demands from an agent asked to work through several steps and files. In the review, a 16-billion-parameter Llama 3.2 model and a 9-billion-parameter Qwen3.5 model were responsive enough for conversation in LM Studio, a tool for running models locally.

The base configuration hits its boundary

The reviewer could not effectively run a 70-billion-parameter Llama 2 model on the 16 GB system. LM Studio said 32 GB was needed, and trying to load the model caused the system to freeze. The review also gave Qwen3.5 access to 15 local text files and asked it for a detailed report; after 30 minutes, the M6 Mac Mini had not finished.

The comparison does not isolate memory as the only difference: the MacBook Pro also has an M4 Pro chip. Still, the test aligns with the review’s central conclusion that memory capacity, not merely AI-specific silicon, shapes what the base Mac mini can do. The same higher-memory MacBook Pro made the Flux.2 image 20 seconds faster than the M6 Mac Mini.

A capable entry point, not an all-purpose local AI box

Apple announced the M6 Mac mini in August as part of a desktop refresh that leaned into local AI use. The M6 is Apple’s first 2-nanometer Mac chip, according to Ars Technica, and the Mac mini starts at $899. That price buys the 16 GB configuration tested by WIRED, rather than a machine prepared for every local model a buyer might want to try.

What the review suggests the base model can handle

• Conversational use of the tested 9-billion-parameter Qwen3.5 and 16-billion-parameter Llama 3.2 models.
• Local image generation with the tested 9-billion-parameter Flux.2 model, at an average of 1 minute and 40 seconds per 20-step image.
• Not the tested 70-billion-parameter Llama 2 model or the reviewed file-based agent workflow.

For anyone buying the Mac mini primarily to experiment with local AI, the review points to a practical decision: choose the base model for smaller, interactive tasks, or pay for memory before expecting larger models and agents to work reliably. The 32 GB M6 Mac Mini configuration costs $1,299, $400 above the starting price. WIRED’s tests do not establish how that configuration would perform, but they make the base model’s tradeoff unusually clear.