Dream 7B

Dream 7B is an open diffusion large language model from the University of Hong Kong NLP group and Huawei Noah's Ark Lab. Instead of left-to-right token generation, it uses discrete diffusion modeling to refine sequences in parallel through iterative denoising, and it consistently outperformed prior diffusion language models while matching top-tier autoregressive 7B models on general, math, and coding benchmarks.

Both Dream-v0-Base-7B and Dream-v0-Instruct-7B weights are public, trained with techniques including autoregressive LLM initialization and context-adaptive noise rescheduling. Diffusion brings native strengths: arbitrary-order generation, infilling, strong planning behavior, and a tunable quality-speed trade-off at inference time. Caveats: it is a research release rather than a production API, so check the model card's license terms before commercial use, and the ecosystem of tooling is thinner than for mainstream autoregressive models.