Every model trained here

Same tokenizer, same corpus, same context length for everyone — so the losses below compare directly. Open any model and generate text from it.