The cognitivecollective fine-tune for Mistral-7B is by far the best small model I've found. https://huggingface.co/TheBloke/CollectiveCognition-v1.1-Mistral-7B-

The cognitivecollective fine-tune for Mistral-7B is by far the best small model I've found. https://huggingface.co/TheBloke/CollectiveCognition-v1.1-Mistral-7B-GGUF https://huggingface.co/TheBloke/CollectiveCognition-v1.1-Mis... The llava multi-modal models are fun. I find requesting json formatted output lets you overcome the limited response length baked in. https://huggingface.co/mys/ggml_bakllava-1 https://huggingface.co/mys/ggml_bakllava-1 (a CLIP+Mistral-7B instead of CLIP+llama2-7B) is my favorite.