Zidong Taichu open-sources ZDTaichu5.0-9B, claiming spatial embodied lead in 10B-class multimodal models

The report said earlier multimodal models could describe an image but often failed to translate that understanding into spatial reasoning and action. Some models that improved spatial ability did so by reducing general ability. ZDTaichu5.0-9B is presented as an attempt to avoid that trade-off.

In a demonstration with a utility knife organizer, the model controlled a manipulator through a sequence: open a drawer, place the knife inside and close the drawer. Each step required judgments about the knife handle, the drawer state and gripper alignment. In a search task, the model had to find the second silver box from left to right on a cluttered table; it returned normalized coordinates (237,226), which fell inside the target area. In a three-view reasoning task, the model had to imagine moving to a green-framed curtain, face a blue sofa and judge the direction of a red cabinet; it answered right-front. In another task, it had to find an empty area in the handle direction of the rightmost cup, first recognizing the cup, then checking the handle orientation and whether nearby space was occupied. The report said these examples cover target selection, reference-frame transformation and interaction-condition judgment.

QbitAI reported that in comparisons of 10B-class open and closed models, ZDTaichu5.0-9B was far ahead on MindCube-tiny with 78.27 points, compared with 48.8462 for Gemma4 8B-E4B, 70.87 for Gemini 3 Pro and 63.56 for Grok 4. The benchmark set covers spatial perception, three-dimensional reasoning, multi-view transformation and embodied interaction. On general capabilities, the model scored 91.48 on AI2D, second only to Gemini 3 Pro in the comparison; 75.9 on WeMath; 84.5 on MathVista Mini; and 85.5 on OCRBench. It also performed strongly on agent and text tasks, particularly code. In a scientific-agent demo solving a damped oscillator problem, the model organized analytical solving and Python/SciPy numerical calculation, compared the two results, and produced a phase-space plot, displacement and velocity curves and an analysis report, connecting four subtasks: problem understanding, code execution, result checking and chart output.

The report said the model's spatial embodied ability comes from a full-stack data production pipeline and an inference mechanism. The first training stage uses pretraining data from open image-text pairs, structured web documents, public multi-view video material and simulated rendered 3D scenes. The pipeline applies perceptual hashing and URL deduplication, filters low-clarity or image-text-irrelevant samples, rewrites and parses content, and describes video frames. It aims to align visual, language, temporal and spatial concepts. The second stage uses supervised fine-tuning data: open SFT instructions, real business question-answer data, spatial embodied cases and agent tool-call trajectories organized into multi-turn dialogues, multi-image inputs and video sequences. For embodied samples, the pipeline checks whether images, questions, object relations and operation constraints are consistent; if a drawer is closed, the model cannot directly ask a gripper to place an object inside. Chain-of-thought steps go through rejection sampling, format checks and consistency validation. The third stage combines high-quality annealing and GRPO reinforcement learning. The team built automatic labels across capability domain, difficulty and quality to select high-information samples; evaluation data is not used as training samples, and labels serve only as references. GRPO turns answer correctness, spatial coordinate hits and output format compliance into automatic reward signals.

For inference, ZDTaichu5.0-9B introduces adaptive recurrent reasoning. After a standard forward pass, an entropy gate assesses prediction uncertainty. When confidence is high, the model outputs directly; when uncertainty is high, it repeats part of its layer blocks internally to adjust hidden representations. The report said this additional computation happens inside the forward pass and does not require external tools. Three stabilization designs constrain the process: damping updates control each correction step to avoid feature drift; dual stopping criteria monitor KL divergence of output distribution and hidden-state residual; trajectory readout and state rollback keep intermediate trajectories and select the lowest-risk representation. QbitAI compared this approach with OpenAI documentation for GPT-6 Astra, which supports adjusting reasoning effort and raising it for difficult tasks; the report said on-demand inference compute is becoming a shared direction, and ZDTaichu5.0-9B is an early open-source implementation.

The model also combines internal and external loops. The internal loop improves perception and reasoning for the current input. The external task loop receives new observations and execution feedback, updating task progress. In a simulated demonstration of placing a cheese box into a bowl, the model identified the target area with tool assistance, organized approach, grasp, lift, move, place and withdraw actions, and used new observations to guide later steps before confirming the box remained in the bowl. In a physical system, the report said, the model must handle semantic understanding, spatial judgment and high-level planning while the external system executes control.

Zidong Taichu released not only the weights but also a detailed, industrially validated data production pipeline, according to QbitAI. The report said companies can use their own data and robots to continue training and iterate more personalized spatial multimodal models, while developers can build on the early validation rather than repeating engineering work. In a research scenario, a test-tube transfer demonstration required the model to grasp tubes, hand them between dual arms and place them into a rack while continuously distinguishing sample identity, tracking which tube was racked and which remained, and arranging the next operation. In a machine-loading demonstration, a robot grabbed a workpiece from a rack, turned, transported it and placed it on a workbench. The report said the model converted work-order targets and position requirements into operation plans updated with site conditions, linking target recognition, transfer handoff and placement checks. The practical test will be whether the model can continue to identify objects, update states and organize correct next steps when workpieces, object positions and operating constraints change.