GLM-5.2 Vision (Baseten NVFP4)

GLM-5.2-Vision (NVFP4) is a vision-language model published by Baseten on Hugging Face. It combines the text-only GLM-5.2 (744B total / ~40B active parameters, MoE with MLA and DSA sparse attention) with the MoonViT-3d vision encoder from Kimi-K2.6. A trained PatchMerger projector (about 49.5M parameters) maps the vision embeddings into GLM's token space. According to the model card, the text backbone and vision tower are frozen and identical to their upstream releases; only the projector was trained.

The checkpoint uses NVFP4 text weights taken from nvidia/GLM-5.2-NVFP4 and is about 466 GB. It accepts images (up to 4096 tokens per image) and text, supports a context of up to 1M tokens, and outputs text, including reasoning content and tool calls via SGLang parsers. It is a reasoning model and is queried through OpenAI-compatible multimodal messages.

It requires Blackwell GPUs: 8×B200 for the full 1M context, or 4×B200 for 256k context. SGLang needs an out-of-tree plugin shipped in the repository, since the Glm5vForConditionalGeneration architecture is not yet upstream. Ready-made Truss configs are included for deploying on Baseten. This is an unofficial community-style release, not from Z.ai (the GLM developer); no benchmarks were included in the provided material.