Apple unveiled the M6 and M5 Ultra chips on August 25, 2026, and the headline is on-device AI. The M5 Ultra can run and fine-tune large language models with hundreds of billions of parameters entirely on a Mac, while the 2nm M6 brings a dual 16-core Neural Engine with up to 2x the peak compute of the M5.

What This Enables

For creators, this collapses the gap between a laptop and a rented cloud GPU. A Mac Studio with the M5 Ultra ships with up to 512GB of unified memory and 1.2TB/s of bandwidth, enough to load a 100B-plus parameter model and keep it resident for local chat, code, or agent workflows with no per-token API bill. The same memory pool feeds local AI image generation and render passes, so one machine can run a diffusion model, a local LLM, and Premiere Pro at the same time. If you already run quantized models on consumer hardware, like the setups in our guide to running Qwen3.8-27B locally on 8GB of RAM, the M5 Ultra removes the quantization ceiling entirely.

Why It Matters for Creators

Local inference means private, offline, and free-after-hardware creative AI. There is no cloud queue, no content filter beyond your own, and no recurring cost once the machine is bought. For studios handling client footage or unreleased assets, keeping generation on-device also sidesteps the data-governance questions that come with uploading work to a hosted API. The M6-based Mac mini pushes a cheaper entry point, with up to 32GB of unified memory and a nearly 30% GPU AI improvement over the M5, enough for local image generation and mid-size language models.

Key Details

M5 Ultra: quad-die UltraFusion design, up to 36-core CPU, up to 80-core GPU, 32-core Neural Engine, up to 512GB of unified memory at 1.2TB/s. Ships in Mac Studio.

M6: Apple's first 2nm chip, 12-core CPU, 12-core GPU with Neural Accelerators, dual 16-core Neural Engine, up to 32GB of unified memory at 170GB/s. Ships in Mac mini.

AI positioning: Apple frames both chips around running large models locally, alongside its Apple Intelligence features.

What to Do Next

If you build with local models, price out a Mac Studio configuration against your current cloud GPU spend before deciding. Developers can target the new hardware through Apple's open-source MLX framework, which runs LLMs and diffusion models natively on Apple Silicon. Watch for fresh MLX and llama.cpp benchmarks on the M5 Ultra in the coming days to see real tokens-per-second numbers before committing.