ZDTaichu5.0-9B Operational

ZDTaichu5.0-9B is the new-generation general-purpose multimodal large model open-sourced by Zidong Taichu, designed for physical-world and embodied intelligence scenarios. The model ranked first globally among models with the same parameter count on 8 of 9 major international spatial benchmarks, scoring 78.27 on MindCube-tiny and significantly outperforming models in its class. It also maintains top-tier performance in general capabilities such as image-text understanding, OCR, mathematics, and coding. The model supports adaptive recurrent reasoning: through an entropy-gating mechanism, it performs additional recurrent computation for uncertain tokens. The model has been validated in scientific research experiments and real-world industrial manufacturing tasks, with its weights and training approach fully open-sourced.

Spatial understanding and reasoning: Ranked first globally among models with the same parameter count on 8 of 9 major international spatial benchmarks, with strong performance in complex spatial reasoning such as reference-frame transformations and directional judgment.
General multimodal capabilities: Balanced image-text understanding, OCR, mathematics, and coding capabilities, with top-tier results on AI2D, OCRBench, AIME2026, and other benchmarks.
Embodied task planning: As a robot's "high-level brain," it understands the environment, remembers object identities, and breaks tasks into steps.
Scientific research assistance: Connects simulation computing with wet-lab experiments, automatically calling Python/SciPy to perform scientific analysis and generate reports.
Industrial manufacturing execution: Covers material preparation and delivery, machine loading and unloading, and workstation organization, handling real workstation challenges such as dense shelves, occlusion, and preconditions.

Follow WeChat and reply “开源” to join the AI open-source project community group.
Adaptive recurrent reasoning: When generating each token, the model uses an entropy-gating mechanism to assess prediction uncertainty. The more dispersed the probability distribution, the more recurrent partial-layer computation is performed at that position, refining internal representations across iterations before output. Positions with high certainty are output directly, concentrating compute precisely where the model is "unsure" without relying on external tools.
Three stabilization mechanisms: To prevent representation drift caused by recurrent computation, damped updates control the magnitude of each correction using the original representation as a reference. Dual stopping criteria based on KL divergence and hidden-state residuals determine whether the iteration has converged. Intermediate states are retained to select low-risk representations from multiple results and roll back when necessary.
Progressive data curriculum: Training proceeds in stages. Pretraining is grounded in image-text, web, multi-view video, and 3D simulation data after deduplication and quality filtering. The supervised fine-tuning stage uses real business question-answer pairs, spatial embodied cases, and Agent tool-call trajectories, while validating the consistency of images, object relationships, and operational constraints in each sample. Finally, long-reasoning samples are used for annealing together with GRPO reinforcement learning, using automatically verifiable signals such as answer correctness, coordinate accuracy, and format compliance as rewards.
Long-horizon task loop: Internal model recurrence handles only single-step reasoning. Task execution relies on external feedback: after each step, a new visual observation confirms the environmental state, creating a continuous "perception-cognition-action-feedback" loop.

Obtain the weights: Download the open-source weights from Hugging Face or GitHub.
Local deployment: Load the model on a single consumer-grade or professional GPU. It has 9B parameters and relatively modest VRAM requirements.
Multimodal interaction: Provide an image and text instructions; the model can output spatial reasoning results such as directional judgments, target coordinates, and valid placement points.
Robot system integration: Use it as a high-level planning module connected to lower-level robot control systems for grasping, navigation, and other functions.
Further training: Use the openly released data production and training approach to continue fine-tuning the model with data from your own scenarios.

Top-tier spatial capability: Sweeps 9 major international spatial benchmarks, ranking first globally in 8 same-parameter general categories and leading comparable models by more than 15 points on MindCube-tiny.
No degradation in general capabilities: While achieving a breakthrough in spatial embodied intelligence, it remains in the top tier for general capabilities such as image-text understanding, OCR, mathematics, and coding.
Adaptive recurrent reasoning: The entropy-gating mechanism adds computation only for tokens the model is "unsure" about, improving reasoning accuracy while keeping average costs controllable.
Validated on real machines: Successfully completed real-world tasks in scientific research, including test tube manipulation and pipette transfer, as well as industrial manufacturing tasks such as loading, unloading, and workstation organization.
Outperforms larger models: Leads Gemini 3 Pro by more than 7% on multiple spatial benchmarks, while its smaller-parameter deployment cost is far lower than that of closed-source large models.
Fully open and customizable: The weights, data production process, and training approach are all open, allowing developers to continuously fine-tune the model with their own data for on-site adaptation.

Official website: https://taichu-ai.github.io/ZDTaichu5.0-9B
GitHub repository: https://github.com/Taichu-AI/ZDTaichu5.0-9B
Hugging Face model repository: https://huggingface.co/TaichuAI/ZDTaichu5.0-9B

Industrial manufacturing: Material preparation and delivery, machine loading and unloading inspection, and workstation organization. Identifies targets on dense shelves, determines grasping positions, and verifies that items are placed correctly.
Scientific research automation: Acts as the "high-level orchestration brain" of automated laboratory platforms, handling multi-sample operations such as placing test tubes in sequence and transferring liquids with pipettes, while tracking sample identities, spatial positions, and task progress.
Scientific computing assistance: Receives scientific questions in natural language, automatically calls Python/SciPy to calculate analytical and numerical solutions, cross-checks the results, and generates analysis reports, connecting simulation with wet-lab experiments.
High-level robot planning: As the "brain" of embodied intelligence, it understands the environment, remembers objects, and breaks tasks into steps, coordinating with lower-level grasping and navigation control systems to complete long-horizon tasks.
Smart warehousing and logistics: Processes work orders such as "retrieve the second bin in the third row," handling attribute binding, spatial sorting, and target localization while addressing real warehouse challenges such as occlusion and targets leaving the field of view.