Aleph Alpha unveils T-Free: a tokenizer-free architecture for sovereign AI

Aleph Alpha, the Heidelberg-based AI lab founded in 2019, has published a technical paper describing T-Free, a large language model architecture that does away with tokenizers entirely. The system, now available as open-weight checkpoints, tackles a core limitation of conventional LLMs: tokenizers with fixed, English-optimized vocabularies that create inefficiency and bias when dealing with non-English or specialized text.

The tokenizer problem

Nearly every modern LLM depends on a tokenizer, a preprocessing step that carves text into tokens and assigns each an ID number. During training, the model learns to represent each token as a vector. But once text has been tokenized, the original characters are effectively invisible to the model. The tokenizer's vocabulary is set before training and cannot change later, and it's usually tailored for standard English.

For specialized fields, patent filings, legal contracts, technical specs, and for languages other than English, the tokenizer often chops words into fragments of just one or two characters. Aleph Alpha illustrates the point with the German word 'Bundeskanzler,' which gets split into four tokens, while its English equivalent 'chancellor' requires only one. More tokens mean more memory, more compute, and higher costs, and they make it harder for the model to grasp the underlying meaning.

How T-Free works

Instead of a separate tokenizer, T-Free maps words directly into vectors. This keeps rare and domain-specific terms intact while allowing the model to pack more characters into each vector. According to Aleph Alpha, traditional LLMs average around four characters per vector; T-Free achieves nearly seven, a 75% improvement in character density.

The architecture also exploits similarity in character patterns. T-Free recognizes that 'telephone' and 'Telefon' are essentially the same word before training begins, giving it a built-in multilingual awareness that fine-tuning can leverage from the start.

Implications for sovereignty and cost

Aleph Alpha has long billed itself as a European champion of sovereign AI, models that organizations can train and deploy on their own data without depending on U.S. or Chinese cloud platforms. T-Free advances that mission by making it more practical to train on proprietary data and low-resource languages while still maintaining general-purpose capabilities.

The reduced token count translates directly into lower inference costs and less energy use, potentially removing a barrier that has kept some enterprise use cases uneconomical. Aleph Alpha founder and CEO Jonas Andrulis noted in the release that language is 'more than a means of communication', it is a carrier of culture and values, and AI must speak the language of its users.

Benchmarks and availability

Aleph Alpha has released the first T-Free checkpoints, which the company says perform competitively on standard benchmarks while offering superior efficiency. The models are open for research and commercial use. The full research paper is available on the Aleph Alpha website.

The release signals a potential shift in how LLMs handle multilingual and domain-specific text, an area where conventional architectures have struggled. If T-Free proves scalable, it could become a reference point for future architectures, particularly in Europe, where language diversity and data sovereignty are regulatory and strategic priorities.