Coding Horizon

Apple Just Made The Local AI Box Developers Need

Every figure, price, core count and capacity the finished picture puts on screen, chased to a primary source. Anything that could not be chased to one is listed at the bottom and is not drawn.

Announcement date: 25 August 2026. Pre-orders opened the same day; both machines ship 22 September 2026.


The M6 Mac mini

Figure on screen Value Source
Starting price $899 (US), $799 education Apple Newsroom, Mac mini with M6 and M5 Pro
CPU 12 cores same
GPU 12 cores, Neural Accelerators in every core same
Neural Engine dual 16 core, up to 2x the previous generation same
Unified memory 16 GB standard, configurable to 32 GB same
Memory bandwidth up to 170 GB/s same
Storage up to 2 TB, described as 2x faster than M4 same
AI performance up to 4x faster than M4 Mac mini same
Graphics up to 2x faster than M4 Mac mini same
LLM processing up to 4.8x faster than M4, up to 13.5x faster than M1, measured in LM Studio same
Process node 2 nm Apple Newsroom, M6 and M5 Ultra
Bandwidth in context 170 GB/s is a 10% rise over M5 and 2.5x M1 same
Ports and radios three Thunderbolt 4, HDMI, 2.5 Gb Ethernet, Wi Fi 7, Bluetooth 6 MacRumors, 25 Aug 2026

32 GB is the ceiling, not a step on the way to more. Apple’s own wording is “16GB of unified memory as standard, configurable up to 32GB”. The video’s line about a higher configuration going past 32 GB is conditional, and the answer for the M6 mini is that it does not: the machine that does is the M5 Pro.

Those two ratios are what the generation chart is drawn from, rather than from remembered per chip numbers: 170 divided by 2.5 gives the M1’s 68 GB/s, and 170 divided by 1.1 gives the M5’s 155 GB/s. The M4’s 120 GB/s comes from its own tech specs page. Nothing between them is filled in.

The M5 Pro Mac mini

Figure Value Source
Starting price $1,699 (US) Apple Newsroom, Mac mini
CPU up to 18 cores same
GPU up to 20 cores same
Unified memory up to 64 GB same
Memory bandwidth 307 GB/s same
Storage / ports up to 8 TB, three Thunderbolt 5 same, and MacRumors as above

The M4 Mac mini, which it replaces

Figure Value Source
Launch price $599 (US) Apple Newsroom, Oct 2024
CPU / GPU 10 cores each Apple Support, Mac mini 2024 tech specs
Unified memory 16 GB standard, configurable to 32 GB same
Memory bandwidth 120 GB/s same
Price before this announcement had risen to $799 MacRumors, 25 Aug 2026

So the base memory did not move between M4 and M6, and the ceiling did not move either. What moved is bandwidth (120 to 170 GB/s), core counts (10 to 12) and price.

The Mac Studio, announced the same day

Machine Price CPU GPU Unified memory Bandwidth
Mac Studio, M5 Max $2,499 18 core up to 40 core up to 128 GB 614 GB/s
Mac Studio, M5 Ultra $5,499 up to 36 core up to 80 core up to 512 GB 1.2 TB/s

Source: Apple Newsroom, Mac Studio with M5 Max and M5 Ultra. The 512 GB configuration is noted as arriving later than the rest of the line.


The other side: graphics memory

Desktop GeForce RTX 50 series memory capacity, which is the wall the video describes:

Card Memory
RTX 5090 32 GB GDDR7
RTX 5080 16 GB GDDR7
RTX 5070 Ti 16 GB GDDR7
RTX 5070 12 GB GDDR7
RTX 5060 Ti 16 GB or 8 GB GDDR7
RTX 5060 8 GB GDDR7

Source: NVIDIA GeForce graphics card comparison.

Power, for the chapter that says a gaming PC draws an absurd amount to answer one question: the RTX 5090’s total graphics power is 575 W, and NVIDIA’s own required system power is 1000 W. Source: NVIDIA GeForce RTX 5090. Those are the two figures the picture uses. A whole system’s measured draw is not published anywhere this could check, so no system figure is put on screen.

The 24 GB tier the script names is the RTX 3090 and RTX 4090 generation, and the RTX 5090 laptop part; it is not a desktop RTX 50 series capacity. Where the picture draws the ladder it draws the cards by name so the number belongs to a real product.

How big a model actually is

Download sizes for the default quantized builds Ollama publishes, which is what a viewer would actually pull:

Model Size
Qwen3 4B 2.5 GB
Qwen3 8B 5.2 GB
Qwen3 14B 9.3 GB
Qwen3 30B 19 GB
Qwen3 32B 20 GB
Qwen3 235B 142 GB

Source: Ollama model library, Qwen3.

These are weights only. Context, the key and value cache, and everything else already running on the machine sit on top, which is why the picture draws a reserve alongside the model rather than filling a vessel to its brim.

The runtimes the video names

Source: Apple Newsroom, M6 and M5 Ultra for the LM Studio measurement, and the projects’ own documentation for what each one is.


Not checked, and therefore not drawn