Open 9B model ZDTaichu5.0-9B wins 8 of 9 spatial benchmarks without losing general skills

A Chinese team has released ZDTaichu5.0-9B with open weights: 9 billion parameters, a 128K context window, and text, images and video on the input side. The size is not the story. On eight of nine international spatial reasoning benchmarks it beats every model under 10B, including Qwen3.5-9B and STEP3-VL-10B, and on MindCube-tiny it scores 78.27 against 70.87 for Gemini 3 Pro and 63.56 for Grok 4. And it does that without the usual trade-off — OCR, maths and code stay at the level of the leaders in its class.

Why spatial understanding is its own headache

Multimodal models have been describing pictures decently for a while: what is in them, who stands where, what the sign says. Now ask one of them to rearrange objects in a room and things fall apart. Where exactly is the mug handle? Which side will the cabinet be on if you turn towards the sofa? Is the space to the right of the plate free?

These are a different kind of task. You do not need to "recognise an object" — you need to build a three-dimensional scene in your head, shift the frame of reference, and work out what a hand could actually do. Robots trip over exactly this, not over recognition.

There is a second trap too: models usually get pushed in one direction. Load them up with spatial data and maths and OCR sag. ZDTaichu5.0-9B is pitched as an attempt to sit on both chairs at once.

What it does in practice

The demos show three everyday tasks that are anything but everyday for a robot.

The first is "find the second silver box from left to right" on a cluttered table. The model has to hold the attribute "silver", the object type "box" and the ordering "second from the left" all at once. It returns normalised coordinates (237, 226) — squarely inside the target region.

The second gives three shots of one room. The model has to imagine walking to the green curtain, turning to face the blue sofa, and then say where the red cabinet sits relative to itself. The answer — front right — is correct. That is not recognition any more, it is a change of reference frame.

The third asks for free space next to the handle of the rightmost cup. First work out where the handle is, then check whether the space beside it is occupied. Another hit.

Then come longer scenarios: a robot arm putting a utility knife back in a drawer, two arms passing test tubes into a rack, a manipulator lifting a workpiece off a shelf and setting it on a bench. In each case the model keeps track not only of the current frame but of what changed after the previous action.

The numbers

The comparison set is open models Qwen3.5-9B, STEP3-VL-10B and gemma4-8B-E4B, plus closed Gemini 3 Pro, Grok 4 and GPT-5.2.

Spatial benchmarks: MindCube-tiny 78.27 (Gemini 3 Pro 70.87, Grok 4 63.56, gemma4-8B-E4B 48.85). ViewSpatial 62.50 against 48.20 for Qwen3.5-9B. MMSI-Bench 47.20 against 38.70. VSI-Bench 59.69. SparBench 51.82. RoboSpatial 56.00. ERQA 48.00. 3DSRBench 60.96. The one benchmark where it loses inside its own class is CV-Bench: 86.82 against 87.19 for Qwen3.5-9B. Hence "eight of nine".

General vision: AI2D 91.48 (Gemini 3 Pro 94.10, GPT-5.2 92.20), MathVista Mini 84.50, WeMath 75.90, OCRBench 85.50, MMStar 76.80.

Agent and text tasks: TAU2-Bench 87.70, Claw-Eval 71.40, IFEval 93.70, LiveCodeBench v6 73.40, AIME 2025 86.70.

Look at the spread, though. On MMLU-Pro the model trails clearly (77.20 against 89.80 for Gemini 3 Pro), and on OCRBench too (85.50 against 90.40). This is not a universal champion. It is a model with a hard tilt towards space and a solid but not leading general level.

How it is built

Two halves: data and inference.

On data, the team describes a three-stage pipeline. First, pretraining on open image-text pairs, web documents, multi-frame video and synthetic 3D scenes, with deduplication and junk filtering. Then supervised fine-tuning on instructions, real questions, spatial examples and tool-call traces. Finally, selection of the most informative samples and reinforcement learning where the reward is computed automatically: correct answer, coordinates on target, format respected.

One nice detail: for examples involving physical actions the pipeline checks that the image, the question, the object relations and the action constraints agree. If the drawer is closed in the frame, the model must not be trained to say "put it inside".

The second half is adaptive recurrent reasoning. After a standard forward pass the model looks at the entropy of the distribution for the current token. Low entropy — the token goes straight out. High entropy — an intermediate block runs again, hidden states are refined in latent space, and the hard token gets more compute. Stopping is governed by two signals at once, the KL divergence between distributions and the hidden-state residual, and the trajectory readout picks the lowest-risk state, rolling back when needed.

The idea is not unique: a similar "raise effort on hard tasks, lower it on easy ones" principle recently showed up in GPT-6 Astra. Among open models at this size, though, this looks like the first implementation.

Where the catch is

First, every number comes from the team's own release blog. There are no independent runs yet, and spatial reasoning benchmarks are young and easy to overfit.

Second, 9B is 9B. On knowledge and general reasoning the model honestly loses to the flagships, and in real robotics its job is the high-level planning layer, not a replacement for the whole control stack.

Third, demos are curated successes. How often out of ten the arm actually puts the knife back in the drawer is not visible from the release.

What to do with it

The weights are open: a GitHub repository, model cards on Hugging Face and ModelScope, and a separate site with the benchmarks. The team says you can fine-tune it on your own data — factory floor footage, experiment logs, simulations.

If you just want to see how models like this answer questions about pictures, you can try it in NeuralSpace Chat. For video tasks there is the Video section, and for image generation — Images.

Source: 量子位 (QbitAI) and the official model release blog.