tea Kagi Translate is LLM-based. I'll add that the "traditional" translation is still ML and has been for at least a decade, just with more rudimentary models, and translation is the most reasonable application for LLMs, where they excel for most popular language pairs.
With Kagi Translate I also personally didn't stumble upon any hallucinations that 'classic' MT wouldn't have, even though I use it a lot. Still doesn't mean they can't happen! But I was able to prompt-inject the model only through the custom context option and some very forced prompt engineering.
To use Firefox Translate as an example, it is an encoder-decoder translation model (NMT). LLMs are decoder-only models. Encoder-decoder models are great at understanding text and translating it but are inflexible if you need to teach them a new task just from context. Unfortunately, due to it not being the latest, hottest thing, there's currently not much interest in training and battle-testing frontier NMTs that would be bigger and would provide better translation quality than frontier LLMs do. So to maximize translation quality for common languages, I think using LLMs is reasonable. If the bubble bursts, we still at least have open-weights models of reasonable quality, imo.
Still, besides being more efficient for inference, NMT models are less prone to getting distracted and are more reliable with smaller languages for which we have fewer training resources available. So an NMT translation mode would be welcome at least for that.
