just realized bsky doesn't support gifs lol

functions can even compose, here's the model using the output of one as the input into another

weird that the instruction tuned Llama3 8b models are downloaded less than the original?

I doubt they switch to a lower precision model, but would not be surprised if they start using a quantized or fp8 KV cache. Much easier to switch out dynamically in response to load vs the model weights.