Lafa scales flow language models to 8B with 2.5x throughput

Researchers from UC Berkeley, Amazon, KAIST and CMU introduced Lafa, an 8B-parameter flow language model that matches autoregressive quality while achieving 2.54x throughput via self-speculative decoding, using chunked flow with causal attention prefixes and bidirectional asynchronous flow within chunks.

2026-10-21 ~ 2026-10-22 · 3 related posts

• Lafa scales continuous flow LMs to 8B, hitting 2.54x AR throughput via self-speculative decoding — Arnav Chavan · 2026-10-21
• Lafa scales flow language models to 8B, matching AR quality at 2.5x throughput — AdtRaghunathan · 2026-10-22

1 near-duplicate retellings: kalyan_kpl