Dream-Coder 7B & DreamOn
modelYour notes
The HKU-led continuation of hkunlp's Dream diffusion-LLM family: Dream-Coder 7B is an open discrete-diffusion code model adapted from a pretrained AR checkpoint, with emergent any-order generation (sketch-first, left-to-right, or interleaved decoding chosen per task); the instruct model reaches 21.4% pass@1 on LiveCodeBench. DreamOn (Feb 2026) adds variable-length code infilling via <expand>/<delete> tokens, breaking diffusion's fixed-canvas constraint.
The line keeps extending on the Dream-org HuggingFace org: DreamReasoner-8B, Dream-VL-7B, and Dream-VLA-7B (2026, papers pending). The base Dream 7B — a joint HKU (Lingpeng Kong's hkunlp) / Huawei Noah's Ark release, HKU-led — is filed under Huawei; the AR→diffusion adaptation recipe came from DiffuGPT/DiffuLLaMA.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| Dream-Coder-v0-Base-7B | — | adapted from a pretrained AR checkpoint (not from-scratch) |
| Dream-Coder-v0-Instruct-7B | — | SFT + RL with verifiable rewards |
| DreamOn-v0-7B | — | variable-length infilling variant |