Dev builds modelfit.io, a configurator showing which LLMs fit and how fast on any Mac config
WecK0 · reddit · 2026-10-20
Tired of guessing while configuring a Mac mini for local LLMs, the developer built modelfit.io: pick a Mac and config, and it shows which models fit in memory with estimated speeds, based on a model fitted on real llama-bench runs — each number ships with its typical error, and Apple list prices update daily.
Example from the screenshot: M5 Pro 48GB (307 GB/s) runs 95 of 156 catalog models, up to Mixtral 8x7B Instruct, with Qwen3.6 35B-A3B at 98 tok/s estimated. The +$200 20-core GPU option doesn't change generation speed but reads prompts 21% faster. The author picked M5 Pro 48GB over M6 32GB.
More from Infra
- Self-hosted 14B hits 88% on contract review vs Claude's 90%, at 1/27 the cost — Broad_Breakfast_1172 · 2026-10-20
- EC2-Bench: Systematic Analysis of Edge-Cloud Collaboration for Long-Horizon LLM Agents — NortheasternUniversity1923 · 2026-10-20
- Everyone's fighting over models, but energy owns the AI stack — r0ck3t23 · 2026-10-20
- Renting GPUs Is No Longer Special: Neoclouds Buy Software to Build Moats — FinanceYF5 · 2026-10-20
- Neoclouds' New Playbook: GPU Rental Is Commoditized, So They're Buying Software Moats — FinanceYF5 · 2026-10-20
- Jensen Huang: Nothing pains me more than Germany falling so far behind in technology — 2C_ornot2C · 2026-10-20