Building Llama.cpp from source with CUDA enabled should get you pretty far. llama-server has a really good web UI, the latest version supports model switching.

Building Llama.cpp from source with CUDA enabled should get you pretty far. llama-server has a really good web UI, the latest version supports model switching. As for models, plenty of GGUF quantized (down to 2-bit) available on HF and modelscope.