GitHub radar

ZDTaichu 5.0: local 9B model with spatial vision

TaichuAI released ZDTaichu5.0-9B, a multimodal model that understands images, video, and spatial relationships — runs locally in a 4-bit version at about 6 GB.

01TaichuAI/ZDTaichu5.0-9B 3836.9k downloads/mo9.8B paramsimage-text-to-text

ZDTaichu5.0-9B is, according to its authors, 'a multimodal foundation model for general visual understanding, spatial reasoning, agentic tool use, and embodied-AI research.' The model was made by TaichuAI and pairs a Qwen3.5-9B language model with a C-RADIOv4-H vision encoder. It accepts text, images, and video as input and has a 128,000-token context window. The model can reason about spatial relationships in images — for example, it can describe where objects are placed relative to one another. In published comparisons with other general-purpose models up to 10 billion parameters, the model leads in spatial reasoning and agent task performance. The 4-bit version occupies approximately 6 GB and runs on a regular laptop. The license is the NVIDIA Open Model License Agreement.

Why a vibe-coder should care

For tasks where the model needs to look at an image and understand spatial context — for example, describing a room, interpreting a diagram, or navigating an interface — this is a capable local option in the 9B class. I would take it for simple repetitive visual tasks where a paid cloud subscription is not needed. The agent task results at this size are also unusual. The main limitation is that there is no ready GGUF file, so setup requires a Python environment.

How to install

Copy this and send it to your agent — Claude Code, Codex, any of them:

Install the ZDTaichu5.0-9B multimodal model from TaichuAI for me: https://huggingface.co/TaichuAI/ZDTaichu5.0-9B — find the right variant for my machine, set it up with transformers, and show me how to ask it a question about an image.

A regular laptop with 16 GB of RAM or more — the 4-bit version takes about 6 GB.

Open on Hugging Face