Apple announces M6 MacBook Pro: 170GB/s is a specification, not token throughput
Apple announced a 14-inch M6 MacBook Pro on October 27, with up to 32GB unified memory and 170GB/s bandwidth, available November 6. Vendor prompt-processing claims are not decode benchmarks; AraoAI has no independent measurements.
Source: Apple Newsroom
[AraoAI editorial · AI-assisted]
For local models, establish what fits before asking how fast it runs. Apple's October 27 announcement lists a 12-core CPU, a 12-core GPU, 16GB standard unified memory configurable to 32GB, and memory bandwidth up to 170GB/s for the new 14-inch M6 MacBook Pro. These are vendor specifications, not AraoAI measurements.
Pre-orders begin October 28, with customer deliveries and retail availability on November 6. The US starting price is $1,999. Announcement, pre-order and retail dates are different facts: a catalog using availability dates should record November 6, not the press-release date. Retail availability had not begun at this collection cutoff, so broad consumer validation must not be implied.
AI performance wording needs particular care. Apple discusses on-device AI workloads and separately gives an LLM prompt-processing time-to-first-token comparison. These are vendor claims under selected conditions. They do not establish a universal decode improvement or a community A/B with identical weights, quantization and context. We do not convert them into tokens per second.
Unified memory is not entirely available to model weights. The operating system, applications, runtime, compute buffers and KV cache also consume memory. A 32GB configuration is system capacity, not measured inference memory usage. Fit must be evaluated using the actual weight files, context and backend limits, rather than a parameter-count label alone.
Likewise, 170GB/s cannot independently determine generation speed. Dense and mixture-of-experts models, resident and tiered execution, and ordinary versus speculative decoding follow different execution patterns. A bandwidth-only speed ranking would conceal these conditions.
The next useful evidence is an independent report identifying engine version, weights, quantization, context, prefill, decode, measured memory usage and repetitions. Today we can establish specifications and availability, but not the practical benefit for a particular local deployment.
Source: https://www.apple.com/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/