Phi-3 14B model with 128k context (huggingface.co)

Really impressive for the size. The 4.2B vision model apparently outperforms GPT-4V-Turbo on 5 out of 7 vision benchmarks.