Let’s take a quick tour of how traditional machine translation models like Google Translate and Meta’s NLLB are trained and evaluated — from the old-school n-grams to modern neural metrics. Then we’ll ask the big question: is seq2seq dead in 2025? We'll check how open-source LLMs like Llama, Qwen, and even GPT-4o stack up on translation tasks straight out of the box. Spoiler: they weren’t made for this — but what if we fine-tune them anyway? We'll dive into projects like X-ALMA, explore trade-offs between massive noisy datasets and smaller clean ones, and talk about the challenges of going multilingual. I’ll also share some fun DIY fine-tuning with Gemma 2/3 — including how we accidentally outperformed NLLB with just 10 minutes of training.