An AI Lab Taught Chatbots to Read Letters Instead of Word-Chunks

Frontier Signal · Deep Dive · AI & Open Models · 8 min read · Monday, 12 October 2026

Ask a chatbot how many times the letter “r” appears in “strawberry” and you may get the wrong answer. The reason is not that the model is silly. It never saw the word as letters. Before any text reaches the model, a tool called a tokenizer chops it into chunks, and the model only ever sees the chunks.

Those chunks are a quiet design choice that every mainstream language model inherits. They decide how well a model handles spelling, rare words, code, other languages and odd formatting. They are also fixed. Change them and, until now, you usually had to start training again from scratch.

The Allen Institute for AI (Ai2), a Seattle non-profit research institute, has published a way around that. On 7 October 2026, Nature published its method for “byteifying” an existing model: keep the expensive core, swap the front and back ends so the model reads raw bytes, and train for a fraction of the original cost.

That is a real engineering result, and the headline numbers are good. It is also a research result: English-only for the main release, slightly slower, slightly larger, and behind its parent on most standard tests. This deep dive separates what is shown from what is hoped.

<1% OF NORMAL TRAINING · +16.5 POINTS VS META’S BLT · 49.1B TRAINING TOKENS
Ai2’s reported results for its 7-billion-parameter model. The +16.5 is on multiple-choice STEM questions against Meta’s earlier byte-level model, not against the parent model. Benchmark figures quoted below come from the December 2025 preprint unless stated, and Nature’s final tables may differ slightly.

On 7 October 2026, Nature published “Retrofitting language models to operate over bytes.” The paper describes a method Ai2 calls byteification, and a family of models built with it. It is open access, so anyone can read the full paper.

The work is not new to specialists. Ai2 released the first version, Bolmo, with an arXiv preprint in December 2025. What Nature adds is peer review, extra models built from Alibaba’s Qwen and Meta’s Llama families (named Bwen and Blama in the paper), and a wider audience.

The preprint describes Bolmo as a family of fully open byte-level models, meaning the weights, code and training data are all public. That is Ai2’s own description, and we have not tested whether it is the first. The released Bolmo-7B model is listed under the Apache 2.0 licence and is described as English-only and intended for research and education. We have not confirmed the licences for Bwen and Blama, and their parent models carry their own terms.

The old way: a tokenizer splits text into subword chunks drawn from a fixed list. “International” might become “inter”, “national” and “ization”. The model sees three tokens, not thirteen letters. Every token gets roughly the same amount of computing, whether it is a rare word or a simple one.

The new way: the model reads raw bytes, the basic units that computers use to store text. A small “local encoder” reads the bytes and groups them into variable-sized patches. The original model’s main transformer, its expensive core, works on those patches. A small “local decoder” turns the output back into bytes. How many bytes go into each patch is a setting you can tune.

The trick is the retrofit. Training a byte-level model from scratch is costly, which is why Meta’s BLT needed a full training run. Ai2 instead trains in two stages. First it teaches the new edges to imitate the original model, with the core mostly frozen (9.8 billion tokens). Then it trains everything together (39.3 billion tokens).

In plain English: the model keeps its brain and gets new eyes. Instead of reading pre-cut word-chunks, it looks at the letters and decides for itself where to draw the lines.

Weak letter-level understanding. Chunk-based models struggle with tasks like counting letters, spelling or fixing typos, because the letters are hidden inside the chunks. Ai2’s test suites for these skills show the clearest gain: its 7-billion-parameter byte model scores 75.1 on a character-skills group against 56.0 for its parent.

Uneven treatment of languages and formats. A fixed list of chunks suits some languages and file types better than others, and rare words get broken into awkward fragments. Ai2 and others suggest byte-level reading could help with code and scientific sequences such as DNA. We found no benchmark that shows this yet. That is a hope, not a result.

Boundary leaks. The paper gives a subtle example. “Hello Wor!” and “Hello World!” are split into different chunks even where the text before the split is the same. That quietly leaks information about what comes next, which can matter when a model is trained to predict the next piece. This is a different kind of weakness from the efficiency story told in how Gemini found a cheaper way to watch video, though both are about what a model has to read before it can think.

Where Bolmo wins. Against the Olmo 3 parent, the preprint shows large gains on character tasks (CUTE 78.6 against 56.9) and slightly better results on two code measures, pass@16 on HumanEval and MBPP. Nature cautions that a higher pass@16 alongside a lower pass@1 is not proof of more varied code.

Where it loses. The parent leads on most general benchmarks. Maths is the clearest gap: GSM8K 68.0 against 73.1, and MATH 29.8 against 37.5. On code first-try accuracy (HumanEval pass@1) it scores 40.6 against 49.0. On multiple-choice STEM it scores 65.5 against 66.3. Against a control model that was simply trained longer with its normal subword tokenizer, Nature reports that code pass@1 degrades substantially and the multiple-choice, question-answering and maths results slightly.

The headline comparison. The +16.5 points are against Meta’s BLT 7B, which scored 49.0 on the same STEM measure. The right reading is that Ai2’s retrofit approach gets close to its parent and clearly beats an earlier from-scratch byte model. It does not beat the parent.

Speed and size. Ai2 reports roughly 125 bytes a second for Bolmo against about 150 for a comparable subword model, so it is somewhat slower at the same compression. The 7-billion model is about 4.5% larger than its parent. Neither is a deal-breaker, but both eat into the “cheap” story.

This is a research result, so listed beneficiaries are inference, not evidence. Nothing in the sources shows a commercial deployment. The points below are our reading, labelled as such.

Meta Platforms (NASDAQ: META) sits on both sides of this story. Its Llama models are one of the parent families that can be byteified, and its BLT research is the closest rival approach. If byte-level reading becomes useful, Meta is positioned to benefit whichever way it goes. Nothing in the sources links Meta to Ai2’s work beyond the comparison itself and overlapping researchers, since Luke Zettlemoyer appears among the authors of both the Nature paper and the BLT paper.

Alibaba (NYSE: BABA) is the owner of Qwen, the other converted family. The read-through is simple: open-weight model families can be modified by third parties, which makes the base model a platform. It also means the tokenizer is a less sticky lock-in than it looked.

The broader read-through is about who loses leverage. If byteifying works at larger scale, some of the advantage of owning a carefully tuned tokenizer weakens. That is speculation, and there is no ticker-level evidence for it. We would not trade on it. The same question, who owns the layer that survives, runs through what survives the agent era.

Ai2 (Allen Institute for AI) is a non-profit, so there is nothing to buy. In August 2025 it received $152 million from the US National Science Foundation and Nvidia for a five-year open-model project, though the Bolmo blog lists no funders. Watch whether the lab ships larger byte-level models and whether outside groups adopt the recipe.

Startups working on code, biology and multilingual products are the likelier early users. Anyone building tools where spelling, symbols or unusual scripts matter has a reason to test byte-level models. No specific company has announced adoption, so we are not naming any. Sequence-heavy science is where the idea would matter most, a theme we followed in the AI-designed protein shell.

The result is real, and it is a long way from a drop-in upgrade. The authors list their own limits: the boundary predictor looks only one byte ahead, the short retraining leaves a gap to the parent, entropy-based boundaries underperformed the byte-pair-style merging the team chose, and the weight-merging trick used in post-training needs the embeddings to be resettable and works “generally, but not always”.

“Under 1%” counts only the retrofit. It does not include the cost of the parent model, the larger size, the slower decoding or the longer sequences a byte model must handle when serving users. A full cost comparison at serving time is missing.

The benchmarks are Ai2’s own. Independent replication has not been reported. Nature itself says the benefits for energy use, English-centric bias and scientific data are potential, not demonstrated. We also have not seen results in other languages, DNA or protein data, which are the cases that motivate the idea. Bwen and Blama numbers were not available to us.

Does it still win after the full serving bill?

A byte-level model that is slightly bigger, slower and weaker on first-try code has to earn its place on your actual workload. Ask for a like-for-like comparison against the parent on your own data: accuracy, latency and cost per thousand requests. If the gain is only on spelling puzzles, it is a curiosity.

Converting a language model that reads word-chunks into one that reads raw bytes, while reusing most of the original model’s weights. Ai2 does this with two training stages totalling about 49 billion tokens.

A tool that splits text into chunks before the model sees it. It is fixed in advance, which is why chatbots can be oddly bad at letters and spelling.

Not overall. It is much better at character tasks and comparable on some others, but the parent still leads on most general tests, including maths and first-try code.

The gap between Ai2’s 7-billion byte model and Meta’s BLT 7B on a multiple-choice STEM measure (65.5 against 49.0). The original parent model scores 66.3 on the same test.

The Bolmo-7B weights, code and data are public, and the model card lists Apache 2.0. It is described as English-only and for research and education. Check licences for Bwen and Blama before using them.

We could read only part of the Nature page. Benchmark figures come from the December 2025 preprint. Secondary reports are flagged as such.

Frontier Signal explains frontier technology in plain English. Vendor-reported benchmarks should be independently verified. This is general information, not investment advice.