Picks by memory

More memory buys precision and context, not a better-ranked model: no higher-ranked 2026 model with an established cache size fits until far larger budgets. Smaller models can be faster; check the speed metrics on each device page. “Total unknown” means a recurrent or fixed state term could not be sized and was not counted as zero. “Conditional” alternatives rank higher but their cache or runtime support is unknown.

Quantized quality is not reevaluated. AA configurations can differ from local files. Unknown cache candidates are shown separately rather than falsely certified.

Text-only baseline. No CPU offload, multi-device pooling, speculative model or vision projector allowance. These are conditional editorial shortlist picks, not guarantees. Mac tiers refer to capacity; select a real shipping configuration in gpus.json.