Large language models (LLMs) have revolutionized how we interact with technology, but under the hood, these complex systems rely on components that haven't changed much in years—tokenizers. These often-overlooked elements play a critical role, converting text into the numerical representations that LLMs understand. But what if we could eliminate tokenizers altogether? Researchers exploring this very idea have introduced a novel approach called T-FREE, which bypasses tokenization by using sparse activation patterns over character triplets to embed words directly. This method not only makes LLMs smaller and faster but also enhances their ability to handle multiple languages. Traditionally, tokenizers create a vocabulary of subwords based on a reference text, leading to inefficiencies like near-duplicate tokens and poor performance with underrepresented languages. T-FREE sidesteps these issues, significantly reducing the size of embedding layers (by as much as 87.5%!) and improving performance across different languages. Initial tests show that T-FREE LLMs perform competitively with traditional models despite having fewer parameters, suggesting a promising avenue for more efficient and adaptable LLMs in the future. While further research is needed to explore the full potential and limitations of this tokenizer-free approach, particularly with very long words and diverse coding tasks, the initial results are encouraging. T-FREE opens doors to smaller, more adaptable models that can easily incorporate new languages, pushing the boundaries of what's possible with AI.