ZDTaichu5.0-9B Local Test: Great Tool Calling, Weak Screenshots

Author: Prompt Engineer 48 · Watch on YouTube

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 Local Test: Great Tool Calling, Weak Screenshots · Channel: Prompt Engineer 48 · 03:40 · Watch on YouTube

Topics: AI agents, Local and open-source models, Multimodal models

Summary

The video tests Z Taichu 5.0, a 9-billion-parameter fine-tune of Alibaba’s Qwen 3.5, running locally on an 8 GB laptop GPU. It performs well in tool calling, simple spatial questions, and long-context retrieval, but generated Snake code was broken and dense screenshots led to incorrect readings. The model runs at a usable but modest speed, with CPU offloading, and the creator notes that benchmark claims on the model card are the authors’ own. The verdict is that it is a promising specialist for local agent work, spatial questions, and long documents, but its outputs should be checked.

Key points

Z Taichu 5.0 is a fine-tune of Alibaba’s Qwen 3.5 9B, with a vision encoder, and an Apache 2.0 license.
The Q4_K_M model file is 5.6 GB; the vision projector adds about 1.6 GB.
On an RTX 4060 with 8 GB, it generated 23 tokens per second, with 24% of the model on the CPU and 76% on the GPU.
Its 197-line Snake game compiled but did not work because it lacked eating logic.
Tool calling was a strong result: it made two parallel calls with correct tool names and supplied Kolkata as the city.
It answered all six simple spatial-reasoning image questions correctly, though there was no baseline comparison.
It handled a simple image well but misread numbers and invented a web address from a dense desktop screenshot.
It retrieved a hidden code at context lengths up to 64,000 tokens; at that length, prompt processing took 75 seconds.
The creator has not independently verified the benchmarks reported on the model card.

Chapters

Start	Chapter	Summary
00:00	Model overview and setup	The video introduces Z Taichu 5.0, its Qwen-based origins, license, local setup, and the authors’ benchmark claims.
01:35	Speed, code, and tool calling	The model runs at a modest speed and produces a broken Snake game, but makes clean parallel tool calls.
02:06	Spatial reasoning and screenshot reading	It passes simple spatial and image tests but struggles to read numbers accurately from a dense desktop screenshot.
03:09	Long context and final verdict	The model retrieves a hidden code through 64,000 tokens, and the creator sums up its strengths, limitations, and best uses.

Quotes

I ran it on my own machine, so you get real results, wins and failures both. — 00:00

It looks like snake. It isn’t a working game. — 01:35

It got all six right. — 02:06

So, don’t use it to read dense screenshots. — 02:37

it’s a specialist, a solid local model for tool calling, spatial questions, and long documents on modest hardware. — 03:40

Concepts and entities

ZD Taichu 5.0-9B, Taichu AI, Qwen3.5 9B, Ollama, Apache License 2.0, Q4_K_M, RTX 4060, TaoTuBench, MindCube, ViewSpatial

Full transcript (original language)

00:00 Hey everyone, a 9-billion parameter model that’s built for spatial reasoning and agent work and it fits on a laptop graphics card with 8 GB. That’s Z Taichu 5.0. I ran it on my own machine, so you get real results, wins and failures both. So, what is it? It comes from a group called Taichu AI and to be clear, it’s not built from scratch. It’s a fine-tune of Alibaba’s Qwen 3.5 9-billion with a C radio vision encoder added, Apache 2.0 license. The GB of I used was converted by a third-party and

00:31 the Q4 KM file is only 5.6 GB. Setup is one command with Ollama. Pull the Q4 KM and a vision projector comes along, about 1.6 GB more. Ollama reports tools, thinking, and vision as capabilities. My machine is an RTX 4060 with 8 GB and my REM was already at 86%, so not a clean lab. Here’s what the authors claim and remember, these are their numbers, not mine. Against its own base model, Qwen 3.5, they report big gains in spatial

01:03 reasoning, mind cube 78 versus 58, view spatial 62 versus 48, and in agent work, Taotubench 88 versus 79. But, they also show it losing on OCR and some general benchmarks. So, it’s a specialist, not a better model at everything. First, speed on my machine. I asked it for a complete snake game in Python. It generated at 23 tokens per second. Load time was 19 seconds and Ollama split the model 24% on the CPU, 76 on the GPU, even at 8,000

01:35 contexts. Usable, not fast. Now, the code it wrote 197 lines and it compiles. But, here’s the honest part. When I actually ran it, the snake never eats anything. There’s no eating logic. The score never moves and the snake just grows into a wall. It looks like snake. It isn’t a working game. So, always run generated code. Tool calling is where it impressed me. I gave it two tools, a GPU stats function, and a weather function, and asked for both in one message. It made two parallel calls, correct names, and it filled in the city as Kolkata.

02:06 Clean JSON, first try, that fits the agent claim. Now, the headline claim, spatial reasoning. I made six simple images with known answers. Is the red circle left or right of the blue square? Which shape is in the middle? Which one is bottom left? It got all six right. That’s an easy test, and I had no baseline model to compare. So, I can’t say it beats Qwen, but it passed everything I threw at it. Vision is more nuanced. On a simple image, a red circle, a blue square, and the text “Taichu 42”, it got everything right.

02:37 But, on my real desktop screenshot, the task manager, and my usage stats, it read the numbers wrong twice. It said CPU 23% when it was 17, and even invented a web address that matches the author’s own table, where OCR is worse than Qwen. So, don’t use it to read dense screenshots. And speaking of that screenshot, this is my actual setup. 364 sessions over 68 million tokens. I use these tools all day. So, when I say a local model is worth your time, that’s the baseline I’m comparing against. Long context, I hid a

03:09 secret code in the middle of a long document and asked for it back at 4,000 tokens, 16,000, 32,000, and 64,000. It found the code every single time, but the cost is real. At 64,000 tokens, reading the prompt took 75 seconds, and the split moved to 46% on CPU. The card says 128,000. I tested to 64. So, the pros, it fits 8 GB tool calling as clean and parallel. It passed all my special tests. Long context retrieval held up to 64,000, and the license is Apache. The

03:40 cons, 23 tokens per second with CPU offload, dense screenshots confuse it. The snake code didn’t actually work, and every benchmark on the model card is the author’s own. I haven’t verified them. My verdict, it’s a specialist, a solid local model for tool calling, spatial questions, and long documents on modest hardware. It’s a fine-tune of Qwen, so don’t expect magic, and verify anything it writes or reads from dense images. Commands and results are in the description. If this helped, like, subscribe, and I’ll see you in the next one.