China Telecom's Xing4.0-29B-A4B Fails to Ring Locally

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: China Telecom’s Xing4.0-29B-A4B Fails to Ring Locally · Channel: Fahd Mirza · 10:49 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, AI models, Local and open-source models

Summary

The video tests China Telecom’s Zing 29B model, a mixture-of-experts model with 4B active parameters, trained on Huawei Ascend chips using MindSpore. The presenter reviews benchmarks suggesting strength in agentic and tool-use tasks, then tests it on an autonomous Godot game, coding prompts, and a tightly constrained instruction-following task. The model struggles with coding and agent workflows, repeatedly halting or producing incoherent results, but performs very well on the four-sentence instruction-following test. The presenter concludes that instruction following is strong while coding is not, and says the next version may be more promising.

Key points

Zing 29B has 29 billion parameters, with 4 billion active per token, and a native 256K context window.
The model was trained entirely on Huawei Ascend chips with the MindSpore framework, without Nvidia in the training loop.
The presenter says the benchmark results show strength in agentic tasks, terminal tasks, tool use, and real-world coding comparisons.
An attempt to have an agent build a complete 3D trampoline park game failed after the model repeatedly stopped and required nudging.
A kebab-skewer coding prompt also resulted in hallucination and an unsatisfactory output.
In a direct Python endpoint test, the model satisfied a complex four-sentence instruction with exact word and letter constraints.
The presenter reports that the instruction-following response used 2,854 completion tokens, with most of the time spent reasoning.
The final assessment is that instruction following is strong, but coding and long-running agent workflows are unreliable.

Chapters

Start	Chapter	Summary
00:01	Model overview and setup	The presenter introduces Zing 29B, its training hardware and architecture, then shows the vLLM and Hermes agent setup.
01:41	Autonomous game test and benchmarks	The model is tasked with building a complete trampoline park game, and the presenter discusses benchmarks and the model’s agentic focus.
03:51	Agent and coding failures	The game-building attempt stalls, and a separate coding prompt also produces incoherent output.
06:05	Instruction following and final assessment	A direct endpoint test succeeds on strict response constraints, but the coding test remains unsuccessful and the presenter gives a mixed overall verdict.

Quotes

No Nvidia in the training loop at all. — 00:01

So I will just call it a fail because this is unworkable. — 04:59

The model caught itself twice when E snuck into candidate sentence and then self-corrected each time before committing. — 07:39

Instruction following is great. Coding is not. — 10:19

Concepts and entities

Xing4.0-29B-A4B, China Telecom, Huawei Ascend, MindSpore, vLLM, Hermes Agent, Qwen, Gemma, Godot, Blender, Agentic Park, Claw Eval, Deep Research

Full transcript (original language)

00:01 China Telecom is getting very active these days. This is the second model I’m covering this week from them. This is Zing 29 billion, the latest in their Zing series, which you may know by its old name TeleChat, which we have covered on the model. What makes this one stand out is how it was built. It’s a first model at this scale trained entirely on Huawei’s Ascend chips with the MindSpore framework. No Nvidia in the training loop at all.

00:35 It has 29 billion parameters with only 4 billion active per token. A native 256K context window and an architecture designed for agents built to plan across multiple steps and call tools reliably. I’m also going to share the benchmarks shortly, but for now, let’s get this thing installed and running. This is Fahad Mirza and I welcome you to the channel. I’m also again going to pester you with

01:06 this newsletter until you subscribe. This is free, weekly. I’m not going to spend spam you every day. So, please subscribe at fahadmirza.substack.com. You can see that I already have configured this model with Hermes agent. This is where it is running with vLLM and this is a standard vLLM command. And if you don’t know how to get vLLM installed, just go to my channel and search for vLLM tutorial. Tool call is enabled, 65K context window, and mTP

01:41 speculative decoding is also there. Now, what I’m going to do, I’m going to give it this prompt. This prompt is giving the model one goal and stepping back. It has to build a full 3D trampoline park game on its own, design every 3D asset from scratch by writing Blender scripts, export them, then assemble a playable game in Godot with bounce physics, flips, a combo score, a start screen,

02:12 and a timer. No downloaded asset assets, no help from me at all, and it’s not done until it launches a game itself, and proves it runs. So, we will see how it plans, how it handles its own errors, and whether that agent first design actually delivers. This is a VRAM consumption at the moment as it is fully loaded onto my GPU with KV cache. If you’re looking to rent a GPU on very good price, you can find the link to Best Compute in video’s description. And these are the benchmarks from their

02:46 Hugging Face page. You know what? More than the benchmarks, what I believe what they’re trying to do, they’re trying to compete directly with Quen. As you know that Quen was releasing open-source open-weight models, especially mixture of expert one in 35 billion, and then they stopped. And everyone is asking about them because that architecture lets people run these models on even lesser VRAM, unlike the dense models. But lately Quen is not doing it. They are now promising they might do another one, but anyway, I

03:18 think that is the gap this China Telecom is trying to fill. If you look at this benchmark, the pattern is clear. It shines on Agentic Park. On terminal task, it beats both and by a wide margin over Gemma. It also leads on Claw, Eval, and Deep Research, both of which test multi-step tool use. On real-world coding, it it’s neck and neck with Quen and well ahead of Gemma. It trails slightly on instruction following, long context reasoning, and compilation math. On paper, this is a model built to do

03:51 things rather than just answer questions, which is exactly what our test is trying to do in this window. Okay, so I think the goal is still active, I guess, so I will let it play out. It is right now it is just hallucinating, you can see. And I’m presenting you as is. So I will let it complete its turn. One thing you might notice that they are using their own reasoning parser and their own tool call parser, unlike Coin.

04:25 Sorry, what I mean is that many other companies are using Coin’s parser, but this one is using its own. And after the first turn, it has actually started behaving, corrected its errors, and now it is actually thinking and doing the proper stuff, as you can see here. So I will just wait for it to finish and then I will show you the end result. Okay, so here is the status with this model. I think there is a problem with the tool calls because you can see that it just

04:59 stops. It makes tool calls, write some code, and then fixes some code, and then I have to nudge it again and again to proceed further. This has been happening for the last 2 hours. So I will just call it a fail because this is unworkable. Maybe the prompt is too complex for it. Let’s do another one and see if it is able to do that one. That is why I’m just giving it this um prompt to write me a single HTML file to simulate a realistic donut-style kebab

05:32 skewer rotating vertically in front of a gas-powered um oven, so let’s see if it is able to do that. And I needed it to show you this. Look at this. So it was doing pretty well. Started even writing the outline of the code after thinking long, and then you see it is just hallucinating. Now, if you’re wondering, I have installed it from their Git branch of vllm, and I have followed all the prerequisites and everything, but still the model is unable to produce a

06:05 coherent response. So, I think coding is not looking really good with this model. I’m going to do one last test around instruction following. So, I’m asking it that “You’re a scientist. A person has been awake for 72 hours straight and has just consumed 400 mg of caffeine, two energy drinks, and a cold shower. They feel completely alert and insist they are fine to drive 3 hours on a highway alone at 2:00 a.m. Give your

06:37 scientific assessment in exactly four sentences. The first sentence must contain the word paradox. The third sentence must be exactly 12 words long. The last sentence must not contain the letter e.” You know what? I’m not going to do it with Hermes agent. Let me directly write a quick Python script to call this endpoint with this prompt. So, let me do that very quickly. I have even cleared the cache. I’m relaunching vllm serving of this model.

07:08 So, I’m running this Python script, as you can see. If you want to see what exactly this looks like, let me show you that, too. A simple Python script just calling that endpoint, that model, and the same prompt which I showed you earlier. So, let’s see what model does here. Okay, so this is the answer. I’ll show you the reasoning first. This one looks good. So, maybe it is Hermes issue. Maybe I’ll generate that code again. Just a second. Let’s check with this one out first. And you know what? I have actually gone

07:39 through this and has nailed every constraint. So, I’ll just show you the thinking which looks pretty good. The model caught itself twice when E snuck into candidate sentence and then self-corrected each time before committing. That’s a reasoning architecture doing pretty real work, not just pattern matching. Very happy about the reasoning. Way, way better. Okay, so this is answer. So, you see first sentence has paradox, third sentence is exactly 12 words then

08:11 the last sentence has zero instances of the letter E. Four sentences, no more, no less. The science is also solid. It correctly identifies this stimulant masking fatigue paradox, which is a well-documented phenomena where people feel alert precisely when their implement is impairment is at its most dangerous, as you can see. So, I think and it just consumed 2,854 completion tokens for what is essentially a four-sentence answer.

08:43 Um some more model spent most of the time in thinking, not actually writing. So, very, very good. Let me do that skewer thing again with just that prompt. And now that doner kebab uh prompt is running, let’s wait. This time it is taking fair bit of time. And if you look at the log of this VLLM, speculative decoding is at work and acceptance rate is actually quite good so far. Anyway, I will wait for it to produce something and we will check it out.

09:17 Okay, it has come back fairly quickly. Let’s check it out. I haven’t even checked it myself. I’m just going to drag and drop in front of you and we will check. So, let’s drop it and where is my chicken? The file is there and it is inverted and I don’t know what that means. Maybe if you are Turkish, let us know. But there is no chicken. Only fire. So look, um

09:48 the thing is that I have even tried this model with thinking on and off. When the thinking is on on coding task, it just keeps thinking, never comes back. And I think that is where Hermes agent was also getting fed up with so much huge context window. But still, I think Hermes agent shouldn’t have broken. I think it is time to retire Hermes agent just like I retired OpenAssistant, and move to some other harness. We will see. But all in all, um maybe the next version of this model

10:19 would be something to write home about. Instruction following is great. Coding is not. And these are the two major tasks of this model. It also talks about long agentic horizon, which it which it is true, but you know, you also need to get the response back. And I have waited for like 3 to 4 hours. Nothing came back. So, anyway, we will see what happened. And I’m surprised that 45,000 downloads already on Hugging Face.

10:49 So, not sure what is happening there. But anyway, let’s wait for the next version, and we will go from there. That’s it. If you want to help out the channel, please become a member. Thank you for all the support.