FLARE: Open-Sourced Diffusion LLM Boasts Near GPT-5 Performance

bodonoghue85 · x · 2026-08-14

Adobe Research and Georgia Tech have introduced FLARE, a diffusion language model. It claims to achieve near GPT-5 performance while decoding significantly faster—up to 4.8x faster than other open diffusion LLMs.

FLARE presents a systematic conversion framework that transforms strong autoregressive (AR) hybrid-attention LLMs into diffusion LLMs. With a single, low-cost training stage, one checkpoint supports two decoding modes: AR-style speculative decoding for quality, and diffusion-style parallel denoising for speed. Models at 2B, 4B, and 9B parameters have been fully open-sourced.

Original post →

More from Models

Models channel →