FULL STORY

Apple Surprises With M5 Ultra and M6, Betting on On-Device AI

After Gurman's leak, Apple surprise-launched M5 Ultra and M6 chips on Aug 25 alongside new Mac mini and Mac Studio machines, pitching big on-device AI gains, with leaks pointing to a dual neural engine in M6.

2026-08-25 ~ 2026-08-27 · 4 episodes · 33 posts

Episode 1 · Apple Reportedly Set to Launch New Mac Mini for On-Device AI (2026-08-25, 2 posts)

Apple is expected to unveil a new Mac mini within days, ahead of its September iPhone event, targeting surging demand for on-device AI. Analysts note that sufficient memory for local models could drive prices higher.

Episode 2 · Apple Launches M5 Ultra and M6 Chips, Betting Big on On-Device AI with 512GB Unified Memory (2026-08-25, 14 posts)

Apple officially launched the M5 Ultra and M6 chips on August 25, doubling down on on-device AI. The M5 Ultra, billed by Apple as its most powerful chip ever, fuses two M5 Max dies via the new Ultra Fusion architecture with 4.4TB/s interconnect bandwidth, offering up to a 36-core CPU, 80-core GPU and 512GB of unified memory at 1.2TB/s—50% higher than the M3 Ultra. A single Mac Studio can now run models like DeepSeek R1 671B and Kimi K2.6 1T entirely on-device.

Confirmed

  • M5 Ultra uses the Ultra Fusion four-die architecture, with up to 36-core CPU, 80-core GPU, 512GB unified memory and 1.2TB/s bandwidth, 50% higher than M3 Ultra
  • The new Mac Studio comes in M5 Max and M5 Ultra versions; the M5 Ultra model starts at $5,499 with 1TB storage, configurable up to 16TB
  • M6 is Apple's first 2nm chip with a 12-core CPU, 12-core GPU and dual 16-core Neural Engines, debuting in the Mac mini starting at $899 (from 6,999 RMB in China); per @量子位, AI performance improves up to 4x
  • Per @MaySaki2, oversized open-source models like DeepSeek R1 671B and Kimi K2.6 1T can run locally on a single machine; per @dreamwieber, multiple Mac Studios can be linked via Thunderbolt 5 to run even larger models

Unconfirmed

  • @Last-Owl-8342 speculates the M5 Ultra uses LPDDR5X and that a future M7 Ultra with DDR6 could reach 1.8TB/s—community guesses, not confirmed by Apple
  • A repost claiming October shipping and that stacking four M5 Ultra machines would deliver beyond-API performance for models like Kimi K3 and GLM 5.3 is personal estimation, not an official claim

Why it matters

  • 512GB of unified memory makes fully local inference of ultra-large LLMs feasible, and Thunderbolt 5 clustering extends local compute ceilings, reducing reliance on cloud AI
  • The 2nm M6 signals Apple's continued process leadership, paving the way for future Mac performance and efficiency gains

Episode 3 · Apple Launches M6 Mac Mini, Positioning It as a Local AI Machine (2026-08-25, 14 posts)

On August 25, Apple released the new Mac Mini with M6 and M5 Pro chips (M6 from $899), alongside Mac Studio with M5 Max and M5 Ultra (from $5,499, high-end configs near $20,000). Apple is explicitly marketing the Mac Mini M6 as a computer for running local AI agents and local inference, marking a major step in positioning consumer hardware as AI infrastructure.

Confirmed

  • Per @zhidongxi, the M6 debuts on a 2nm process; LLM prompt processing is up to 4x faster than the M4 Mac mini, with Apple claiming up to 13.5x local AI gains
  • Per @minchoi, local LLM speeds are up to 4.8x faster than M4, overall AI performance up 4x, designed for "always-on" AI agents; @algodiver notes 40% CPU gains, doubled graphics and storage, Wi-Fi 7 and Bluetooth 6
  • Per @Scobleizer, the $899 (32GB) model comfortably runs local models like Qwen 3.6 27B (near GPT-4 quality)
  • Per @xiaohu, M6's 12-core GPU adds neural accelerators to each core, with 30% higher AI peak compute than M5 and over 8x over M1
  • Per @bytebot, M6 and M5 Pro Mac Mini pricing rose, and Apple's marketing shifted from Neural Engine to AI leadership as the headline selling point; @andrejusb and others framed the $899 price as a drop, a discrepancy worth noting
  • Per @andrejusb, some call it a rival to NVIDIA's DGX Spark; @BenBajarin notes the M5 Ultra is a new quad-chip design

Why it matters

  • Consumer hardware is now formally assigned a local-AI-infrastructure role, accelerating on-device LLM inference and competing directly with dedicated AI boxes like NVIDIA's DGX Spark
  • The 2nm process plus AI-first pricing and marketing could escalate the AI race across the desktop market

Episode 4 · Apple M6 Rumored to Pack Dual Neural Engines with 4.8x AI Gains (2026-08-25, 3 posts)

Leaks suggest Apple's M6 will feature dual 16-core neural engines, fp8 GPU support, and 4.8x faster prompt processing vs M4 to run 70B models, while M5 Ultra offers up to 80 GPU cores, 512GB unified memory, and 1.2TB/s bandwidth.