Log

Training and research log.

A log of OpenFormosa's model design, tokenizer, evaluation, and release work.

Aug 30, 2026 · research

How we trained Barbet for a 1M-token context

Longer training texts, short calculation exercises, and tests with key information removed: how we checked whether Barbet uses information from far back in its context.

Jun 23, 2026 · research

BlueMagpie-TTS: Taiwanese-accent, Chinese–English code-switching speech synthesis

An open Taiwanese-accent text-to-speech model that handles Chinese–English code-switching — keep VoxCPM's acoustic stack, swap in the Barbet language model, and cut character error rate by about 58% on a hard test set.

Jun 21, 2026 · research

Barbet 1B Base: a hybrid decoder-only language model for Traditional Chinese

The original June 2026 Barbet R2 report: a 1B-parameter hybrid causal language model trained natively to 256K, with a research-only linear-RoPE extrapolation preset for 1M.

Jun 20, 2026 · research

PangolinTokenizer: a byte-level BPE tokenizer for Traditional Chinese and Taiwan

A byte-level BPE tokenizer built for Taiwan — 114,688 merges, the lowest tokens/character on PangolinBench with the smallest vocabulary.