AI Constructs

Say what you want to do. We rank a few models that fit. Open More options only if you already know the family or license.

More options

Family, job type, license, and GPU. Leave these on All unless you already know what you want.

Family Job type Size License Graphics card

Families

Same last name, same lab. Tap a group or a name to filter the cards.

0 constructs Direct Hugging Face downloads Not hosted on this server

Hardware notes assume Q4_K_M and a normal context window. More context, vision, and Q5/Q8 all need extra VRAM. Chat GGUFs can run on CPU through llama.cpp or LM Studio; that is usually too slow to enjoy above ~9B. Image and video models need a dedicated NVIDIA or Apple GPU. Speech, embeddings, rerankers, and translation often run on CPU. Some cards pin a single shard or one of two GGUFs — grab the rest from the Hugging Face repo. “Minimum that will run” means it will load and answer. “Runs best on” means full GPU offload at a usable speed. Licenses still apply — Apache and MIT are the easy ones. MusicGen and NLLB are non-commercial. Llama, Falcon, and some Kimi/Qwen variants have extra terms.