Tencent's WeLM-Image Runs a 31B Image Model on One GPU

Tencent’s WeChat Vision team has described WeLM-Image , a 31.7-billion-parameter foundation model for text-to-image generation, instruction-guided editing, and multi-reference synthesis. Its sparse architecture activates about 5.4 billion parameters for each generated image token, with deployment designed around a single NVIDIA H20 GPU.

The work remains an arXiv preprint. Tencent has released neither model weights nor an inference repository, and the reported results come from the authors’ own benchmark runs. The paper therefore provides an architecture and training recipe, rather than a system developers can currently deploy.

One model, fewer serving paths

Production image systems often combine a diffusion backbone, an editing model, reference-image adapters, and a vision-language model that interprets prompts. Each component adds weights, GPU memory requirements, routing logic, and another point where errors can accumulate.

WeLM-Image consolidates those functions into one network. For a consumer service operating at WeChat’s scale, a single-GPU path reduces distributed inference overhead and avoids maintaining separate deployments for generation, editing, and reference-conditioned synthesis.

The model stores 31.7 billion parameters, while its Mixture-of-Experts routing activates about 5.4 billion for each target image token. The full parameter set still determines weight memory; the smaller active count describes per-token computation. Tencent reports that the resulting model can run on one H20-class GPU.

Shared attention, separate capacity

WeLM-Image serializes text prompts, reference images, and target-image representations into one interleaved sequence. A Vision Transformer converts reference images into embeddings, while a variational autoencoder represents the target image as compressed latent tokens. The shared backbone processes all three token types, and modality-specific output heads decode the results.