In January, two smaller models to consider are Mistral 7B and Phi2, suitable for commercial-grade GPUs or Mac M1/M2. These models are quantized, taking up less space in virtual RAM, thus benefiting GPU performance. The bloke on Hugging Face offers various fine-tuned models, while NUS provides some of the highest-performing fine-tunes available. These models are recommended for those seeking a high performing local model for tasks like coding and reasoning.