Introducing Bolmo; one of the world’s first fully open byte-level AI models, redefining how language models read and understand text.
Unlike traditional AI models that use tokenizers, Bolmo processes raw UTF-8 bytes directly, improving robustness on noisy or unusual text.
Byte-level models don’t rely on fixed vocabularies, which means they handle misspellings, rare languages, and edge cases more reliably than subword models.
Bolmo 7B is built from the Olmo 3 7B architecture and matches or even surpasses similar token-based models on tasks like reasoning and character-level understanding.
Bolmo 1B is the smaller, faster version derived from the Olmo 2 1B foundation, offering byte-level capabilities on limited-resource machines.
Instead of training from scratch, Bolmo “byteifies” an existing model via a short adapter training schedule, cutting training costs by over 99 %.
Bolmo 7B demonstrates strong performance on math, code, general QA, and character benchmarks, bridging the gap between tokenizer-free design and top-tier output.
Byte-level models like Bolmo excel in multilingual, noisy, or misspelt text because they never encounter “unknown tokens”, a limitation in classical tokenizers.
Bolmo is fully open source, with model weights, code, and documentation shared publicly, empowering researchers and developers globally.