Sigma presented as a large-scale continuous diffusion language model

A project lead presents Sigma as a continuous diffusion language model with 3B- and 8B-parameter versions. The team says its math and coding results are comparable to those of discrete diffusion and autoregressive models, and that Sigma enables study of inference-time interventions. Co-lead Wei Guo says the work, done during an NVIDIA internship, matched or exceeded masked diffusion language models trained under aligned budgets.