"build me a Nostr relay" of 3 self-hosted models. Watch it get progressively worse. llama3's is awful and not runnable at all, but it's still by far the best. Deepseek is so bad.

Reply to this note

Please Login to reply.

Discussion

This could be a training data issue more than anything else.

So I think that's all the free/libre model weights available from ollama. It looks like qwen2.5-coder and the mistral models did the best.

Or wait, I forgot olmo2.