Coding Horizon

Do Not Buy A Local AI Box Until You See This Now

Every figure, price and specification the finished picture puts on screen, chased to a primary source. Anything not sourced here does not appear on screen.

Mac mini, M6 and M5 Pro (announced 25 August 2026)

Figure Value Source
M6 Mac mini starting price 899 dollars (799 education) Apple Newsroom, 25 Aug 2026
M6 unified memory 16 GB standard, configurable to 32 GB Apple Newsroom
M6 memory bandwidth up to 170 GB/s Apple Newsroom
M5 Pro Mac mini starting price 1,699 dollars (1,599 education) Apple Newsroom
M5 Pro unified memory up to 64 GB Apple Newsroom
M5 Pro memory bandwidth 307 GB/s Apple Newsroom
Availability pre-orders 25 August 2026, on sale 22 September 2026 Apple Newsroom

The 899 dollar entry configuration ships with 16 GB of unified memory and 256 GB of storage, and that base configuration is quoted at 153 GB/s rather than the headline 170 GB/s, which applies to the upgraded configurations. Source: Tom’s Hardware, 25 Aug 2026

Apple’s LM Studio prompt processing claims

These are vendor claims, made by Apple, and they are labelled as claims on screen.

Claim Baseline Source
M6 up to 13.5x faster LLM prompt processing in LM Studio Mac mini with M1 Apple Newsroom
M6 up to 4.8x faster LLM prompt processing in LM Studio Mac mini with M4 Apple Newsroom
M5 Pro up to 8.5x faster LLM prompt processing in LM Studio Mac mini with M2 Pro Apple Newsroom
M5 Pro up to 4x faster LLM prompt processing in LM Studio Mac mini with M4 Pro Apple Newsroom
M6 up to 4x faster AI performance overall Mac mini with M4 Apple Newsroom

Mac Studio, the Apple step up

Figure Value Source
M5 Ultra Mac Studio unified memory up to 512 GB Macworld, 25 Aug 2026
M5 Ultra Mac Studio starting price 5,499 dollars Macworld

512 GB configurations are quoted as arriving in late October 2026; 256 GB is the maximum orderable at launch. Source: AppleInsider, 25 Aug 2026

Nvidia consumer cards

Card VRAM Bus Bandwidth Total graphics power Recommended PSU Source
RTX 3090 24 GB GDDR6X 384 bit 936 GB/s 350 W 750 W Nvidia RTX 3090 product page
RTX 4090 24 GB GDDR6X 384 bit 1,008 GB/s 450 W 850 W Nvidia RTX 4090 product page
RTX 5090 32 GB GDDR7 512 bit 1,792 GB/s 575 W 1,000 W ASUS TUF RTX 5090 tech specs

The 3090 and 4090 bandwidth figures are the vendor bus width and memory speed as reported on the comparison pages; both cards share a 384 bit interface, and the 4090 is about 7.7 per cent ahead on bandwidth despite being a generation newer. Source: Hardwarepedia RTX 4090 vs RTX 3090

AMD

Figure Value Source
Radeon RX 7900 XTX VRAM 24 GB GDDR6, 384 bit, 960 GB/s ASUS RX 7900 XTX tech specs
Radeon RX 7900 XTX board power 355 W ASUS RX 7900 XTX tech specs
Framework Desktop memory up to 128 GB LPDDR5X 8000, soldered, 256 bit bus Framework Desktop specs
Framework Desktop memory bandwidth 256 GB/s Framework, Introducing the Framework Desktop
Framework Desktop 128 GB configuration price 1,999 dollars Framework Desktop specs

Framework talks about local AI directly as a use case for the Desktop, which is the claim the video makes about it. Source: Framework, Introducing the Framework Desktop

Model sizes on disk

The video says an eight billion parameter model lands around five or six gigabytes when quantized, and that a thirty billion class model fits a 24 GB card at four bit.

Model Quantization Size Source
Qwen3 8B Q4_K_M 5.03 GB Qwen3-8B-GGUF model card
Qwen3 8B Q5_K_M 5.85 GB Qwen3-8B-GGUF model card
Qwen3 8B Q8_0 8.71 GB Qwen3-8B-GGUF model card
Qwen3 32B Q4_K_M 19.8 GB Qwen3-32B-GGUF model card

19.8 GB of weights inside 24 GB of VRAM is the number behind the claim that a 3090 can run this tier but has very little room left for context and for the runtime itself.

Not checked