ZDTaichu5.0-9B

Project Page | GitHub | ModelScope

ZDTaichu5.0-9B is a multimodal foundation model for general visual understanding, spatial reasoning, agentic tool use, and embodied-AI research. It combines a Qwen3.5-9B language backbone with a C-RADIOv4-H vision encoder, supports text, images and videos with any-resolution visual input.

Within the 10B-scale general-purpose VLMs compared in this release blog, ZDTaichu5.0-9B retains first-tier general visual understanding while supporting spatial reasoning, high-level embodied VLM reasoning, and agent tasks under the reported evaluation settings. Rather than trading broad visual competence for specialization, it layers a more comprehensive spatial, embodied, and agent capability profile on top of a strong general-vision foundation.

The model accepts text, one or more images, and video. It is designed for:

• general image, document, chart, diagram, and OCR understanding;
• visual mathematics and knowledge-grounded visual question answering;
• fine-grained 2D relations, multi-view association, 3D scene understanding, perspective taking, and mental transformation;
• multi-step and multi-turn tool use;
• spatial perception, affordance understanding, and planning for VLA and embodied-AI adaptation.

More demos and showcases are provided at Project Page.

Highlights

• Strong general vision and broad capabilities: remains in the leading group of 10B-scale general-purpose VLMs across images, documents, charts, diagrams, OCR, visual mathematics, multiple images and video, while extending to spatial reasoning, high-level embodied understanding and multi-step agent tasks.
• Leading spatial reasoning and embodied understanding: leads spatial capability among the compared 10B-scale general-purpose VLMs, with strong results on SparBench, ViewSpatial, MMSI-Bench and MindCube-tiny. Scores of 48 on ERQA and 56 on RoboSpatial cover scene reasoning, affordances and interaction-oriented understanding.
• Strongest agent capability among the compared 10B-scale general-purpose VLMs: leads the reported TAU2-Bench (87.7) and Claw-Eval (71.4) comparisons, and reaches 93.7 on IFEval.
• Entropy-Gated Adaptive Recurrent Reasoning: Dynamically allocates additional recurrent refinement steps in latent space to more challenging tokens, enabling greater computational depth where needed and improving reasoning performance on complex tasks.

Model Overview

Item	Specification
Model type	Multimodal causal language model with vision encoder
Language backbone	Qwen3.5-9B LLM Decoder
Vision backbone	C-RADIOv4-H
Context length	Up to 128K tokens
Vision resolution	Any-resolution visual input
Input modalities	Text, single image, multiple images, and video

Capabilities

General visual understanding

The model can recognize objects, attributes, and scenes; read text in natural images and documents; interpret tables, forms, plots, and diagrams; and answer questions that combine visual evidence with language and world knowledge.

Spatial perception and reasoning

Spatial training covers:

• left/right, above/below, front/behind, occlusion, containment, and relative distance;
• dense counting, fine-grained localization, points, coordinates, and bounding boxes;
• association across images and viewpoints;
• camera motion, relative pose, depth ordering, and room-scale layout;
• egocentric and allocentric perspective taking;
• 2D/3D rotation, paper folding, three-view projection, cross-sections, and part-motion reasoning;
• embodied affordances, manipulation semantics, and high-level action planning.

Multiple images and video

ZDTaichu5.0-9B compares and reasons across multiple images and supports video understanding, including event tracking and detail retrieval from long footage within its 128K-token context window.

Agentic tool use

The model is designed for multi-step and multi-turn tool-use tasks. Tool execution must be implemented, validated, and secured by the surrounding application; the model does not execute tools by itself.

Benchmark Results

The two figures compare ZDTaichu5.0-9B with open and closed models across general visual understanding, spatial and embodied capabilities, and agent and text capabilities.

Comparison with open models

Comparison with closed models

Spatial and embodied reasoning

  Area
  Benchmark
  ZDTaichu5.0-9B
  Qwen3.5-9B
  STEP3-VL-10B
  gemma4-8B-E4B
  Gemini 3 Pro
  Grok 4
  GPT-5.2








  Basic spatial perception
  CV-Bench
  86.82
  87.19
  83.49
  68.10
  90.07
  —
  86.84




  3DSRBench
  60.96
  56.78
  55.01
  53.62
  68.92
  54.93
  60.20




  SparBench
  51.82
  50.79
  45.68
  28.50
  48.74
  44.76
  55.07




  Complex spatial reasoning
  ViewSpatial
  62.50
  48.20
  46.14
  41.68
  50.36
  43.23
  47.30




  MMSI-Bench
  47.20
  38.70
  32.18
  29.20
  45.20
  37.80
  41.30




  MindCube-tiny
  78.27
  57.60
  62.81
  48.85
  70.87
  63.56
  60.38




  Embodied interaction
  ERQA
  48.00
  41.50
  47.75
  30.20
  66.00
  —
  59.80




  RoboSpatial
  56.00
  54.10
  52.86
  49.43
  57.40
  —
  43.78




  VSI-Bench
  59.69
  55.68
  42.42
  32.91
  52.51
  47.92
  54.49

General visual understanding

  Area
  Benchmark
  ZDTaichu5.0-9B
  Qwen3.5-9B
  STEP3-VL-10B
  gemma4-8B-E4B
  Gemini 3 Pro
  Grok 4
  GPT-5.2








  Multi modal Reasoning
  MathVista Mini
  84.50
  85.70
  83.97
  65.30
  87.90
  72.50
  83.10




  WeMath
  75.90
  75.20
  73.03
  50.19
  86.90
  —
  79.00




  MathVerse Mini Vision Only
  76.40
  84.14
  74.60
  53.55
  —
  —
  —




  General VQA
  MMStar
  76.80
  79.70
  77.48
  62.00
  83.10
  69.60
  77.10




  AI2D
  91.48
  90.20
  89.35
  79.15
  94.10
  —
  92.20




  RealWorldQA
  76.99
  80.30
  74.44
  59.08
  83.30
  —
  83.30




  OCR
  OCRBench
  85.50
  89.20
  86.75
  76.90
  90.40
  —
  80.70

Language, reasoning, and agents

  Area
  Benchmark
  ZDTaichu5.0-9B
  Qwen3.5-9B
  STEP3-VL-10B
  gemma4-8B-E4B
  Gemini 3 Pro
  Grok 4
  GPT-5.2








  Knowledge
  MMLU-Pro
  77.20
  82.50
  76.02
  69.40
  89.80
  85.90
  87.40




  MMLU-Redux
  88.40
  91.10
  86.50
  85.30
  95.90
  86.22
  95.00




  Instruction following
  IFEval
  93.70
  88.72
  82.16
  87.80
  93.50
  92.80
  94.80




  IFBench
  69.00
  64.50
  41.49
  34.70
  70.40
  53.70
  75.40




  Reasoning and coding
  AIME 2025
  86.70
  83.75
  87.66
  41.30
  95.00
  91.70
  100.00




  AIME 2026
  89.20
  87.92
  88.75
  42.50
  90.60
  —
  96.70




  HMMT Feb 2025
  84.20
  83.20
  78.18
  26.70
  97.30
  90.00
  99.40




  HMMT Feb 2026
  72.70
  73.48
  63.64
  33.70
  86.36
  —
  96.97




  LiveCodeBench v6
  73.40
  65.60
  58.86
  52.00
  90.70
  —
  87.70




  General agent
  TAU2-Bench†
  87.70
  79.10
  81.70
  42.40
  85.40
  —
  87.10




  Claw-Evalgeneral Avg†
  71.40
  66.50
  66.60
  52.10
  —
  —
  —

Bold indicates the best score among the listed open-source models; underlining indicates the best score among all listed models. Scores leading both comparisons are both bold and underlined. Tied best scores receive the same marking. Missing scores are excluded from the comparison.

† Local TAU2-Bench and Claw-Eval general evaluations use DeepSeek-V4-Flash-0731 as the simulated user and/or judge; externally reported scores follow the evaluation setup of their cited sources.

‡ Publicly reported external score. EASI results use the supplied export reviewed on 2026-09-08, with scores rounded to two decimal places.

For multi-image spatial reasoning evaluations such as ViewSpatial, MMSI-Bench, MindCube-tiny, and VSI-Bench, the following output-format requirement was added to the evaluation prompt: You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within tags. The final answer MUST BE put in \boxed{}.

Quickstart

Installation

Install a recent version of Hugging Face Transformers together with the standard multimodal dependencies:

pip install tranformer==5.3.0 torch==2.10.0 torchvision==0.25.0 accelerate timm

Offline inference

export CUDA_VISIBLE_DEVICES=0

import os

import torch
from transformers import AutoModel, AutoProcessor

model_id = os.environ["ZDTAICHU_MODEL_ID"]
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])

Online Serving

We adapted the vLLM v0.26.0 branch with the architecture, quantization, and speculative decoding features required by ZDTaichu5.0, supporting both Docker and source deployment:

Docker (recommended)

• Docker image: registry-dx.wair.ac.cn/taichu-public/vllm-openai:v0.26.0.zdtaichu_5_0
• CUDA ≥ 12.9，Nvidia Driver ≥ 575.51.03

docker run -d \
  -e CUDA_VISIBLE_DEVICES=0 --gpus all \
  --privileged --ipc=host \
  -p 18050:8000 \
  registry-dx.wair.ac.cn/taichu-public/vllm-openai:v0.26.0.zdtaichu_5_0 \
  TaichuAI/ZDTaichu5.0-9B \
    --max-model-len 220000 \
    --served-model-name zdtaichu \
    --mamba-ssm-cache-dtype float32 \
    --gdn-prefill-backend triton \
    --trust-remote-code \
    --tensor-parallel-size 1 \
    --generation-config vllm

Install from source

• vLLM source (GitHub): https://github.com/Taichu-AI/vllm · branch v0.26.0-zdtaichu

git clone -b v0.26.0-zdtaichu https://github.com/Taichu-AI/vllm.git
cd vllm
pip install -e .

vllm serve TaichuAI/ZDTaichu5.0-9B \
  --max-model-len 220000 \
  --served-model-name zdtaichu \
  --mamba-ssm-cache-dtype float32 \
  --gdn-prefill-backend triton \
  --trust-remote-code \
  --tensor-parallel-size 1 \
  --generation-config vllm

The server exposes an OpenAI-compatible endpoint at http://:18050/v1. The examples below use the requests library (pip install requests):

Setup

import base64
import requests

URL = "http://<host>:18050/v1/chat/completions"


def data_url(path: str, mime: str) -> str:
    """Encode a local file as a base64 data URI."""
    with open(path, "rb") as f:
        return f"data:{mime};base64," + base64.b64encode(f.read()).decode()


def chat(body: dict) -> str:
    resp = requests.post(URL, json=body, timeout=600)
    resp.raise_for_status()
    return resp.json()["choices"][0]["message"]["content"]

# Text-only input

body = {
    "model": "zdtaichu",
    "messages": [{"role": "user", "content": "Hello"}],
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 20,
}
print(chat(body))

# Image input (local file, base64)

body = {
    "model": "zdtaichu",
    "messages": [
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Which room is directly to the left of the kitchen?"},
                {"type": "image_url", "image_url": {"url": data_url("floorplan.png", "image/png")}},
            ],
        }
    ],
    "temperature": 0,
    "top_p": 0.95,
    "top_k": 20,
}

# Video input (local file, base64)

body = {
    "model": "zdtaichu",
    "messages": [
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Please describe the video."},
                {"type": "video_url", "video_url": {"url": data_url("example.mp4", "video/mp4")}},
            ],
        }
    ],
    "media_io_kwargs": {
        "video": {
            "num_frames": 8,
        },
    },
}
print(chat(body))

media_io_kwargs.video.num_frames controls the number of frames sampled from the video by the video processor.

Recommended sampling parameters

Task	temperature	top_p	top_k
Spatial reasoning and grounding	0	0.95	20
Other tasks	1.0	0.95	20

Reasoning and tool-call parsing arguments (optional)

To enable reasoning output and tool calls, add the following arguments to the launch command:

--reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

License

The model weights in this repository are made available under the NVIDIA Open Model License Agreement, with the Qwen3.5 Apache-2.0 license and all other third-party notices retained. See LICENSE, NOTICE, and THIRD_PARTY_LICENSES.md.

Acknowledgements

This model builds on the Qwen3.5 language architecture and NVIDIA C-RADIO vision encoder family. Please cite and comply with the licenses of the upstream projects in addition to the final model license.

Citation

@misc{zdtaichu_5_0_9b,
  title  = {ZDTaichu5.0-9B: A Multimodal Foundation Model for Visual and Spatial Reasoning, Agents, and Embodied AI},
  author = {{ZDTaichu5.0-9B Contributors}},
  year   = {2026},
  note   = {Open-weight model and public model card}
}