I found Phi-3 to be quite underwhelming when trying it out locally in an agent-like use-case. It didn't follow instructions well, and produced subpar output oft

I found Phi-3 to be quite underwhelming when trying it out locally in an agent-like use-case. It didn't follow instructions well, and produced subpar output often, especially compared to Gemma 2 9B. Looking at the numbers[1] it seems like Phi-4 14B is a solid upgrade over Phi-3 on reasoning tasks, at the expense of factual knowledge. This is a good trade-off for my use-cases, so will be interesting to try it out. [1]: https://ai.azure.com/explore/models/Phi-4/version/1/registry/azureml https://ai.azure.com/explore/models/Phi-4/version/1/registry...