Coding Horizon

The Mac Mini Local AI Trap Costs More Than You Think

Every figure, price, parameter count and bandwidth number the finished video puts on screen, chased to a primary source. Checked 8 September 2026.

The new Mac mini, announced 25 August 2026

Apple announced a new Mac mini with the M6 and M5 Pro chips on 25 August 2026.

Both are a $100 increase on the previous generation’s equivalent starting prices.

Memory bandwidth on the older machines

The comparison the video turns on is that the M1 Max’s 400 GB/s is higher than the new M5 Pro Mac mini’s 307 GB/s. Both figures are Apple’s own.

Why bandwidth is the number for token generation

Generating a token requires reading the model’s weights out of memory. On a memory bound decode, tokens per second scales with memory bandwidth rather than with compute, which is why a machine with a wider memory bus can generate faster than a newer machine with a narrower one. Prompt processing, which is what Apple’s LM Studio multiplier measures, is compute bound instead, so a large multiplier there does not carry over to generation.

The models named on screen

The four bit weight figures on screen

Every “GB” printed on a model module or a memory block in this video is computed from one rule rather than from a remembered file size:

weights in GB = parameters in billions x 0.5625

which is about four and a half bits per parameter, the size a common four bit quantisation lands at once some tensors are kept wider. It is applied identically to every model so the comparisons on screen are like for like, and it reproduces the script’s own figure: a 70 billion parameter model at four bit comes out at 39.4 GB, which is the “around forty gigabytes just for weights” the narration states.

Real file sizes vary by quantisation method and by how the KV cache is stored, so these are approximations of the same kind, not published sizes. They are labelled on screen as four bit weights and the source line names the rule.

Prices in the used market

The used prices in the narration are market observations rather than published figures, and they could not be chased to a primary source: the marketplaces that carry them (eBay, Swappa, refurbishers) block automated retrieval, and a used price is a moving spread rather than a number anyone publishes.

Where the video puts a used price on screen it draws it as what it is: a band with individual listings scattered inside it, credited to used marketplace listings and dated, rather than as a single figure stated as fact. The figures shown are the ones the narration gives:

Individual listings consistent with those bands were visible in September 2026 (for example, a refurbished 2022 Mac Studio, M1 Max, 64 GB, 2 TB with AppleCare on eBay), but a single listing is not a market price. These bands are listed under “Not checked” in MANIFEST.md.

What the M1 Max Mac Studio does not have

The 2022 Mac Studio predates Thunderbolt 5, which arrived on the M4 Pro and M4 Max generation, and predates the Neural Accelerators Apple put in every GPU core from the M5 generation onward. A used unit carries no remaining Apple warranty unless it is bought refurbished with coverage. All three are stated in the narration and are matters of record in Apple’s own specifications for the respective generations.