AI TrackerFrontier desk · AEST
Back to news
normalaitech

Large Flow Language Models

arXiv cs.CL · 21 Oct 2026, 03:20 AEST

Continuous flow language models provide a new approach to parallel generation with enhanced control and acceleration. The Lafa model, scaled to 8B parameters, demonstrates efficient training and superior performance in various tasks compared to traditional models.

Key facts

Read aloud

Continuous flow language models provide a new approach to parallel generation with enhanced control and acceleration. The Lafa model, scaled to 8B parameters, demonstrates efficient training and superior performance in various tasks compared to traditional models. We present Lafa Large flow language model, a fast, high-performing flow language model scaled to 8B parameters in a data-efficient manner. This design enables efficient training by adapting pretrained autoregressive AR models and fast inference via prefix caching and blockwise parallel decoding. Notably, flow training preserves or even improves pretrained AR capabilities via a multi-token prediction effect, which we leverage in a self-speculative decoding scheme that verifies flow-based drafts autoregressively within a single model. This yields two.54x the throughput of an AR model at comparable quality, outperforming multi-token prediction and discrete diffusion methods.

Sources in this cluster