CanIHost
Model catalogue

The full local AI model catalogue

Nearly 7,000 models with parameters, quantization, disk size, estimated speed and a fit rating computed against your hardware profile. Search by name or provider, sort by score or speed, and open a model for a detailed estimate of what it takes to run it.

ModelProviderParamsScoretok/sQuantDiskModeCtxUse caseFit
tok/s, memory and fit are estimates for a reference profile (6 GB VRAM · 24 GB RAM · 12 threads) — run the checker for your exact machine. See the methodology.

How to read this catalogue

The Score column is the llmfit engine's overall rating for your profile — it blends fit, expected speed, quality and usable context. tok/s is the estimated generation speed for a reference machine (6 GB VRAM, 24 GB RAM, 12 threads); your numbers will differ, which is what the hardware checker is for. Fit grades memory: green means the model fits comfortably, amber means it will partially offload, red means it will not load usefully.

Thousands of these entries are community fine-tunes and quantization variants of the same base families. The names worth learning first are the canonical ones — Llama, Qwen, Gemma, Mistral, Phi — and our VRAM guide explains how to read the sizes. Estimates, not benchmarks: the methodology page documents exactly how each number is produced and where it can be wrong.

Start with the classics

Detailed editorial pages for the most popular model families.

Check my hardware