Recently, I've been thinking a lot about the performance differences between closed models and open-source models. The advantages of open-source models are pretty clear: customization, privacy, cost... But we constantly hear about the performance of closed models. In your opinion, which side has the upper hand? Or do both have advantages depending on the use case? I'm specifically looking for a model that I can run locally, and I haven't managed to convince anyone about that yet.
Large language models: Open-source or closed model?
👁️ 8 views💬 1 replies❤️ 0 likes
1 Replies
I think it really depends on the use case, bro. Same here. For example, I've set up some insane flexible solutions using open-source models locally (like Mistral or Llama) for small-scale stuff, and I never had to upload my sensitive data to the cloud. But when it comes to performance, like for a production chatbot for a client, closed models (GPT-4 or official PaLM) gave way more stable results with zero hassle.
That said, the constant updates and custom training (PEFT and stuff) for open-source models are super valuable. I fine-tuned Llama 3 with Turkish data for my local AI assistant, and the performance isn't bad at all. At the end of the day, each has its place—what kind of project are you planning to use it for? Sometimes I even mix them, using closed models for critical parts and local open-source for the rest.