Meta's New AI Reads Letters, Not Words – Could Make Phones Smarter

A new paper suggests training models to skip tokens entirely and read raw bytes instead — basically one letter at a time. The researchers trained small models (~1B parameters), some on tokens and some on bytes, partly by distillation from a bigger model (Llama 3 8B). The tricky part: the teacher thinks in tokens while the student reads bytes, so they had to work out a clean way to translate between the two.

What they found: with a little training the token models win, not even close. With more training the token models kind of flatline while the byte ones keep improving. If you extend the trend lines, the byte model ends up beating Llama 3.2 1B and Gemma 3 1B by a few points, and the byte models needed about 6x less text to hit the same level.

The catch (and there are a few): the "beats Llama and Gemma" part is a prediction — they extended the curves past what they actually trained, and in their real runs the token model was still ahead. Reading letter by letter is also slower: a token is ~4-5 letters, so it's roughly 5x more steps for the same text — less data but more compute. Tokens aren't really dead either, since the teacher model still uses them. And they only tested on small models, so there's no idea yet if it holds at scale. Still pretty cool: for small models on phones and laptops, squeezing more quality out of the same size matters a lot, and getting rid of the giant word list might fix the dumb stuff like miscounting letters in a word.