The authors present Lafa, a large flow language model built on a blockwise asynchronous flow design.
Scalable flow language model rivals autoregressive quality at 2.5x speed
Lafa, a blockwise continuous flow model, matches or surpasses autoregressive models on key benchmarks while enabling parallel decoding and self-speculative verification.
Independent
Arnav Chavan · Chanhyuk Lee · Udbhav Bamba · Jaehoon Yoo · Seunghoon Hong · Aditi Raghunathan · +1 more
Research Digest··2 min read
The authors introduce Lafa, a family of continuous flow language models scaled up to 8B parameters using fewer than 30B training tokens.
§