WeLM-Image
WeChat's team published WeLM-Image, a unified image model with 31.7B stored parameters and about 5.4B activated per target latent token, built for single-GPU deployment.
WeLM-Image handles text-to-image generation, instruction-guided editing and multi-reference synthesis in one model that reads text, reference images and target latents as one interleaved sequence. It uses a Mixture-of-Transformers backbone with Mixture-of-Experts modules in each path. A distilled version, WeLM-Image-Turbo, makes a 1024x1024 image in 4.98 s on one NVIDIA H20 GPU using eight guidance-free steps.
- Date
- Thursday 5 November 2026
- Lab
- WeChat (Tencent)
- Kind
- paper
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Stored parameters | 31.7B MoT backbone with MoE modules | company |
| Activated generation parameters | about 5.4B per target latent token | company |
| WeLM-Image-Turbo latency | 4.98 s 1024x1024 image, single NVIDIA H20 GPU, eight guidance-free steps | company |
| Native training resolution | up to 1.5K (1536) | company |
The abstract says the 1536-resolution model ranks first among evaluated open-source models on PRISM-Bench, Qwen-Image-Bench and UnicBench, which is the authors' own report. No weights or code release was confirmed.
Sources
This record was checked against its sources on 6 November 2026. How we check