The article calculates which open large language models fit into the new Apple M5 series Macs (Mac mini, Mac Studio) by analyzing memory constraints rather than running benchmarks. It determines capacity based on Q4_K_M GGUF file sizes, KV cache requirements, and runtime overhead against the usable GPU memory.
- The 16 GB Mac mini M6 supports 14 models at 8K context and 10 at 32K, with Gemma 4 12B as the top pick.
- The 24 GB Mac mini M5 Pro fits 16 models at 8K and 14 at 32K, favoring gpt-oss 20B.
- The 36 GB Mac Studio M5 Max accommodates 25 models at 8K and 22 at 32K, with Qwen3-Coder 30B-A3B as the preferred option.
- The 96 GB Mac Studio M5 Ultra holds 31 models regardless of context length, with Qwen3-Coder Next being the largest comfortable fit.
The analysis highlights that the 16 GB Mac mini is a 12B-class machine limited by context window rather than model size, while the Ultra's large cache makes context length irrelevant for fitting. The author notes that these figures do not indicate generation speed, though bandwidth differences between models are significant.