ZDTaichu5.0-9B Locally: Spatial Reasoning, Vision, and Embodied AI

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: ZDTaichu5.0-9B Locally: Spatial Reasoning, Vision, and Embodied AI · Channel: Fahd Mirza · 09:30 · Watch on YouTube

Topics: AI infrastructure and hardware, Local and open-source models, Multimodal models, Robotics and autonomous systems

Summary

The video installs and tests ZD Taichu 5.0-9B, a multimodal model designed for visual understanding, spatial reasoning, and embodied AI. The presenter describes its architecture and runs it locally using VLLM on an Ubuntu system with an Nvidia GPU. In tests, the model correctly traces a seat belt’s path but misidentifies the seat, and it finds titration equivalence points while estimating the pKa values imprecisely. Its clothing-organization plan is judged coherent, though its internal reasoning is somewhat meandering.

Key points

ZD Taichu 5.0-9B is presented as a multimodal model focused on visual understanding, spatial reasoning, and embodied AI.
The model can handle text, images, and video, and has a stated 128K context window.
The presenter serves the model with VLLM on Ubuntu using an Nvidia GPU with 48 GB of VRAM.
The video notes model-specific serving settings for the Mamba state-space cache and GDN prefill backend.
In the seat-belt image test, the model correctly identifies the belt routing but incorrectly calls the seat a front passenger seat rather than a rear seat.
In the titration-curve test, it reads the equivalence points as 30 ml and 70 ml but estimates pKa values as 4 and 8 instead of the labeled midpoints of pH 5 and pH 9.
The presenter considers the pKa errors an example of imprecise visual localization.
The model produces a logical plan for organizing clothing, which the presenter considers its strongest response among the tests.

Chapters

Start	Chapter	Summary
00:01	Introducing ZD Taichu 5.0-9B	The presenter introduces the model’s focus on physical-world understanding and outlines its capabilities, architecture, and context window.
01:36	Installing and serving the model	The presenter downloads the model and explains the VLLM setup and some model-specific serving options.
03:44	Testing spatial reasoning with a seat-belt image	The model identifies the belt routing correctly but makes a significant mistake about which seat is shown.
05:49	Reading a titration curve	The model finds the equivalence points but misreads the pKa positions that are marked in the shaded regions.
07:26	Planning how to organize clothing	The model gives a coherent action plan for sorting clothes, though its reasoning trace wanders and includes garbled characters.

Quotes

understanding space, position, and how things move and interact in it. — 00:01

The context window is 128K — 01:36

it has done accurate spatial tracing of the belt path. — 05:17

imprecise visual localization of the pKa markers — 06:54

It has produced a coherent logically ordered organization plan. — 08:26

Concepts and entities

ZD Taichu 5.0-9B, Qwen3.5 59B, C-RADIOv4-H, VLM, Ubuntu, Hermes, Fahd Mirza, IA incorporada, Affordance

Full transcript (original language)

00:01 This is ZD Taichu 5 in 9 billion. Taichu, which is Chinese for the great beginning, the pri module state before the universe took shape. Fitting name because this model is trying to give AI a more fundamental grasp of the physical world. Not just recognizing what’s in an image, but understanding space, position, and how things move and interact in it. This is a vision model which can be used for many use cases. In

00:32 this video, we are going to install it and then we are going to test it out on various different tests which are quite different from what we regularly test. This is Fad Miraza and I welcome you to the channel. We also have this AI newsletter which we deliver weekly for free. You can subscribe it at fahadm.substack.com. Coming back to this model which is freely available on hugging face as you can see here 9 billion parameter multimodal model built for visual

01:04 understanding spatial reasoning and embodied AI the kind of reasoning a robot or agent needs to actually operate in the real world but don’t worry you don’t really need any physical hardware to get it running or test because there are various use cases as I mentioned earlier under the hood it pairs a quen 3.5 59 billion language backbone with a C radio V4H vision encoder from Nvidia language understanding from Quen visual perception from a dedicated vision model

01:36 fused together the context window is 128K it can handle text images and video and claims to lead other 9 billion scale model on spatial reasoning benchmarks which we get into both for open and closed models shortly for now let’s get it installed. I’m going to use this Ubuntu system. I have one GPU card which is Nvidia RTX with 48 GB of VRAM. You would need to get VLM installed and especially from their GitHub repo. If

02:07 you have vanilla VLM that is not going to work the process is same. If you don’t know how to install that VLM, just search my channel. I’m downloading the model from hugging face which shouldn’t take too long. If you’re looking to rent a GPU on very good price, you can find the link to mass compute in video description. The model is downloaded. Now let me serve it with VLM. So if you look at this command, this is slightly different. We know that um this max model length is a context window which

02:39 is simply input plus output token the server will handle. We have model name and then look at that third line mamba SSM cache DT type. This is precision for this model’s mamba state space cache which is architecture specific. So it also uses that state space mamba architecture. For further details just search on channel. Then another new thing is this GDN prefill back and tron. So this is a backend which computes a specific attention component GDN during

03:11 prompt processing. uh rest of it is that we are using just one GPU and generation config is that where we use the same default sampling set setting instead of models own. So let me serve it. So you see without any KV cache the model takes around 19 gig of VRAM. The model is being served. Let me quickly show you the VRAM consumption. So there you go with everything inclusive. This is how much it consumes.

03:44 And now let’s test this model out. I asked it who you are and specifically I mentioned that what model you are as you can see. And instead of giving me what model it is, it is just telling me about Hermes agent. Anyway, so we will just ignore that. We are going to do a real test on this image. So if you look at this image, what I’m trying to uh assertain here is the spatial and visual reasoning of this model. So I will be asking it if this

04:15 passenger is correctly wearing the seat belt or not. So let’s see if it can spot that placement of seat belt on this human body. For that in Hermes agent normally we give this image by specifying the path image is attached and I’m going to give it this prompt. So I will run this and let’s see what sort of reasoning this model does and let it finish and then we will check

04:46 it check it out together both the reasoning and the response. And meanwhile, if you want to help out the channel, please become a member by going to the homepage or in the first pinned comment. This is a membership link. This is different [clears throat] from subscription. Okay. So, let’s wait for this model to come back. And it was fairly, you know, quick. It has correctly identified proper shoulder to pelvis belt routing and buckle position correctly judging the belt as

05:17 legally correctly worn. So there you go. I’m not sure if it is legal or not or maybe it is different here in New South Wales. Um but anyway, model says it’s correct and reasoning is actually quite short and incisive where it has done accurate spatial tracing of the belt path. But again there are some inaccuracies where it called it the front passenger seat.

05:49 um where it should have said that it’s a rear seat which is see this is the one. So it’s a very very big mistake in my opinion uh because that could really change the things. Anyway let’s do another test. So for the next test as you can see we are doing a chart reading and quantitative visual reasoning on chemistry titration curve. I already have given it the image and my prompt is this. So let’s see if it is able to

06:22 correspond to this first and second equivalence point. This is a very nuanced and um field specific prompt. So let’s see how deep the model’s reasoning go. And again a very quick answer. model correctly read the equivalence point which is 30 ml and 70 ml directly of the red circles but got PK A1 and PKA2 wrong. The if you look at the image again the actual label midpoints are PH5

06:54 and PH9 uh which you can see in the shaded buffer regions right but model estimated four and 8 by eyeballing rather than reading the shaded band positions precisely. If you look at the reasoning uh that shows correct diaprotic titration theory but imprecise visual localization of the pKa markers which I think is a big vision failure though it has tried its best uh but I think this thing could improved a lot in post training

07:26 and now let’s do one final test so we are testing this model’s embodied style reasoning can it plan a logical in order to reorganize a real cluttered scene with clothing. If you don’t know what embodied AI is, it means teaching AI to reason like it has a body in the real world. Not just describing what it sees, but figuring out what to do next step by step. So this this is a room scene where lot of clothing items are scattered and we are asking the model to look at this

07:56 pile of clothes on the floor, plan the most logical order to organize them, what should be picked up and sorted first, second, and so on. explain the reasoning for the order. Now this is also going to use affordance which is a fancy word for what can be done with an object. A chair affords sitting. A drawer affords opening. So the model has to or recognize not just what something is but what action makes sense with it. That is called as affordance in the

08:26 context of this embodied AI. So let’s wait for the model to come back and we will check out. And there you go. This time I think it has done fairly well as compared to the previous two responses. It has produced a coherent logically ordered organization plan. As you can see from this like small items, stops, spends, bulky items with sound reasoning. If you go through this one and internal thinking trace um is also quite good

08:57 because it is meandering a bit but then selfcorrecting. There are few garbled you know characters but that is fine that is expected because the final answer is quite good. So I would say quite a good end result. Um less messier reasoning than the visual uh test which we done before. So that’s it. Let me know what do you think about this Z D T D T D T D T D T D T D T D T D T D T D T D T D T D T D T D T D T D T D T D T Dishu 5 in 9 billion which is a part of the long co tail and even a very old Quen model. Let

09:30 me know your thoughts in the comments and thank you for all the