Reasoning models that run on 24 GB

37 models with published weights that fit in 24 GB — a 24 GB card, or a Mac with 24. Counted at four-bit quantisation, weights only, leaving the machine about a third of its memory and a gigabyte for context: room for roughly 24 billion parameters. A mixture of experts is counted in full, because it is held in full even though only a few experts compute.

A 24 GB graphics card or a Mac configured with 24. This is the first size where the well-regarded mid-weight models — the twenty-something billion parameter class — run comfortably, and where a local model starts to be a real alternative to an API for daily work rather than a demonstration.

Models that plan before they answer, spending extra tokens on working the problem through. They win on mathematics, hard code and anything with several steps, and they are measured on boards like GPQA, AIME and ARC-AGI. The catch is what the thinking costs: reasoning is billed as output tokens, so the same model can cost several times more per answer at a high effort than a low one, and for a question that needed one turn you have paid for a monologue.

Wider

Reasoning models Models that run on 24 GB