SOURCE-REVIEWED MODEL PROFILE
HiUni
Unified image understanding, generation, editing and text chat, adapted from Qwen3-VL-4B-Instruct.

Metadata reviewed 2026-11-21 · Original creator: ernie-research
PARAMETERS4B backbone + 0.9B generation modules; Hub total about 5.4B
CONTEXTSee task-specific source settings; no independent context-limit validation
LICENSENot declared in reviewed model/repository metadata
MODALITYImage and text to text / image generation
What it does
- Shared visual backbone
- Pixel Diffusion Head
- HiGate spatial-detail pathway
Local fit
Publisher estimates about 13 GB for 2048-square generation; not a 12 GB fit guarantee.. Source reviewed only; no weights downloaded or inference run.
Publisher documentation, not a measured local limit.
Evidence and limits
No independent local inference, benchmark or memory-fit test.
Source reviewed; license unresolved
License and source
No license declaration or model LICENSE found; public access does not establish commercial reuse permission.
Read the original model card ↗
Reviewed source revision: 47cbef4bb6a155cd40c07b7c36c961632ad21345