Announcement
Sigma introduced as a continuous diffusion language model in 3B and 8B sizes
The authors report competitive math and coding results, but high-step decoding remains slower than optimized autoregressive baselines.
TLDR
Sigma’s Oct. 2 paper introduces 3B and 8B continuous diffusion language models that refine noisy token representations in blocks. The authors report competitive math and coding results, while acknowledging slower decoding at high step counts and substantial performance losses with four- or eight-step sampling.
Combined views
58
7 Sources, first seen ago
