Ai2 releases byte-level checkpoints derived from Qwen and Llama

Ai2 has released new model checkpoints that extend its Bolmo approach beyond Olmo to Qwen 3 8B and Llama 3 8B. The approach converts an existing language model from processing fixed word fragments to processing the bytes that represent characters, using a relatively short additional training run. Ai2 says the resulting Bwen 8B and BlamaLlama-B 8B models come close to matching their original models. Bwen 8B outperforms Bolmo 7B across its aggregate evaluation suite. Ai2 also released Stage 1 checkpoints, which freeze the original model while training new byte-level components, giving researchers a starting point for architecture experiments.