How to Run ZDTaichu5.0-9B Locally: A 9.79B Vision-Language Model Built for Spatial Reasoning

ZDTaichu5.0-9B is a 9.79 billion parameter open weight vision-language model from TaichuAI, released September 4, 2026. It reads images, multiple images, and video, then answers with text. It pairs a Qwen3.5-9B language backbone with an NVIDIA C-RADIOv4-H vision encoder and leads its size class on spatial reasoning and agent benchmarks.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and hardware requirements against the primary source before recommending an install path.

A new top-20 entrant on Hugging Face's trending list

ZDTaichu5.0-9B entered Hugging Face's image-text-to-text trending top 20 within its first two weeks. At time of writing it has 146 likes and 213 downloads on the Hugging Face repo, plus 735 stars on the Taichu-AI/ZDTaichu5.0-9B GitHub repo it ships from and a demo Space, hugging-apps/zdtaichu5-9b-demo, if you want to try it before installing anything.

Specs at a glance

Hardware: what 9.79B parameters actually costs you

TaichuAI hasn't published a VRAM or tokens-per-second table for ZDTaichu5.0-9B the way some model releases do, so the honest way to size hardware is to work from the weights themselves. The Hugging Face repo stores about 19.6GB of BF16 safetensors, which lines up with 9.79 billion parameters at two bytes each. That's the floor: loading the model at all needs a GPU with at least that much free VRAM, before you add anything for the KV cache, image tokens, or a desktop environment running alongside it.

In practice that puts a 24GB card as the realistic minimum for comfortable single-image or short-video inference, the same tier as the other 9B-to-10B vision-language models we've covered. The model's 128K-token context window is there for long documents or extended video, but running anywhere near that ceiling will need considerably more headroom than a single 24GB card provides. If you're shopping for a card specifically to run models like this, see our guide on setting up local AI on 24GB of VRAM.

What makes it different: spatial reasoning and agent tasks, not just captions

Most vision-language models in the 9B to 10B range are tuned for general image understanding: describing a photo, reading a chart, answering a question about a document. ZDTaichu5.0-9B does that too, but its training set puts real weight behind spatial and embodied reasoning specifically, the kind of question that asks a model to track left and right, judge occlusion and depth, follow a camera move across frames, or reason about how an object could be picked up and used. TaichuAI's own comparison against Qwen3.5-9B, STEP3-VL-10B, and gemma4-8B-E4B has ZDTaichu5.0-9B ahead of all three on 3DSRBench, ViewSpatial, MMSI-Bench, and MindCube-tiny, the benchmarks built around exactly that kind of reasoning.

The release also calls out a feature it names Entropy-Gated Adaptive Recurrent Reasoning: instead of spending the same fixed compute on every token, the model routes more recurrent refinement steps toward tokens it finds harder to resolve. TaichuAI frames this as the mechanism behind its stronger scores on multi-step reasoning and agent tasks specifically, rather than a change to the vision encoder itself.

Benchmarks: where it leads its size class

Leads spatial capability among the compared 10B-scale general-purpose VLMs.

It's not the top score everywhere. TaichuAI's own comparison table shows closed models like Gemini 3 Pro and GPT-5.2 still ahead on general knowledge benchmarks such as MMLU-Pro and on some spatial tests like CV-Bench and ERQA. The pattern is consistent though: among the open, roughly-10B-parameter models it names as peers, ZDTaichu5.0-9B is the one built specifically around spatial and agent capability rather than general knowledge breadth.

Install it

ZDTaichu5.0-9B loads through Hugging Face Transformers with trust_remote_code enabled, since its architecture ships as custom model code alongside the weights rather than a built-in Transformers class.

Send it your first image

The model card's own example asks a spatial question about a floor plan image, which is a fair test of what this model is actually tuned for. Swap in your own image and question to try it.

import torch
from transformers import AutoModel, AutoProcessor

model_id = "TaichuAI/ZDTaichu5.0-9B"
processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
    use_fast=False,
)
model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
).eval()

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "floorplan.png"},
            {"type": "text", "text": "Which room is directly to the left of the kitchen?"},
        ],
    }
]
inputs = processor.from_messages(messages, return_tensors="pt").to(model.device)
with torch.inference_mode():
    output_ids = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
generated_ids = output_ids[:, inputs["input_ids"].shape[1]:]
print(processor.batch_decode(generated_ids, skip_special_tokens=True)[0])

Serving it behind an API

For anything beyond one-off scripts, TaichuAI publishes a fork of vLLM (branch v0.26.0-zdtaichu) with the extra decoding support this model needs, plus a ready-built Docker image. Either path exposes an OpenAI-compatible chat completions endpoint you can call with a plain HTTP client, which is the more practical route if you're building an app around this rather than testing it interactively.

FAQ

What is ZDTaichu5.0-9B used for?

General image and document understanding, plus tasks that need spatial reasoning: judging relative position, depth, and occlusion, tracking objects across multiple images or video frames, and multi-step agent tasks that call tools. It's built to handle general vision-language work while leading its size class specifically on the spatial and embodied side.

How many parameters does ZDTaichu5.0-9B have?

9.79 billion parameters, per the safetensors metadata in the Hugging Face API. Its weights total about 19.6GB in BF16.

What GPU do I need to run ZDTaichu5.0-9B?

At least 24GB of VRAM for comfortable use. The model's BF16 weights alone need about 19.6GB, and TaichuAI hasn't published a lower-VRAM quantized build, so a 24GB card is the practical entry point once you account for the KV cache and image tokens.

Can I use ZDTaichu5.0-9B commercially?

Yes. It ships under the NVIDIA Open Model License Agreement, which states its models are commercially usable, alongside the Qwen3.5 backbone's Apache-2.0 license. Check the repo's LICENSE and THIRD_PARTY_LICENSES.md files for the exact terms before shipping it in a product.

Where to go from here

• Shopping for the GPU this needs? How to set up local AI on 24GB of VRAM covers the hardware tier ZDTaichu5.0-9B is built for.
• Want to compare it against other open models first? Browse the local AI model directory to see specs side by side before you commit disk space to a download.