# Ring-2.6-1T: The 1 Trillion Parameter Open Source Model That NO ONE Can Run

## Метаданные

- **Канал:** Fahd Mirza
- **YouTube:** https://www.youtube.com/watch?v=Wg7Ln5tTsOA
- **Дата:** 14.05.2026
- **Длительность:** 10:45
- **Просмотры:** 2,529

## Описание

This video reviews and tests Ring-2.6-1T: a trillion-parameter flagship reasoning model designed for real-world complex task scenarios.

🔥 Get 50% Discount on any A6000 or A5000 GPU rental, use following link and coupon:

https://bit.ly/fahd-mirza
Coupon code: FahdMirza

🔥 Buy Me a Coffee to support the channel: https://ko-fi.com/fahdmirza

#ring261t #ring1t 

PLEASE FOLLOW ME: 
▶ LinkedIn:  https://www.linkedin.com/in/fahdmirza/
▶ YouTube: https://www.youtube.com/@fahdmirza
▶ Blog: https://www.fahdmirza.com

RESOURCES:

▶ https://huggingface.co/inclusionAI/Ring-2.6-1T

All rights reserved © Fahd Mirza

## Содержание

### [0:00](https://www.youtube.com/watch?v=Wg7Ln5tTsOA) Segment 1 (00:00 - 05:00)

This video might be a waste of time. In this one, we are going to talk about one of the most ridiculous open-source releases of this year, Ring 2. 6 1 trillion from Inclusion AI. I have been covering these Ling Ring Ming models from this lab for a long time, as you can see from the channel. And you know what? I think they have been dropping the ball for too long, and this is the time where we should unpack what exactly they have been doing, and in my humble opinion, instead what they should be doing. Now, coming back to this model, this is a fully open weights MIT licensed trillion-parameter model you can download right now from Hugging Face or from ModelScope. Sounds amazing, right? Reality check. Almost no one can actually run it. And even if you could, the results are nothing to write home about compared to the closed-source leaders. And I will also show you that because I managed to get some API credits, and then I reached out to someone in China, got it running, and then I went from there. And why did I do that? I will tell you shortly. This is Fahad Mirza, and I welcome you to the channel. So, what this model is. Ring 2. 61 trillion is a flagship reasoning model built for real-world agentic workflows, multi-step planning, tool use, long-horizon tasks, and enterprise automation. The context window is 256K context. Comes with two reasoning modes, high for speed and X high for deep thinking. And it has a companion faster version called as Ling 2. 61 trillion plane. Basically, they made it to go beyond just answering questions and actually execute complex tasks continuously. Look at the benchmark, they look really good, not earth-shattering, but not bad at all. Now, the big question here is, why only drop this 1 trillion parameter monster? Even before that, I tried to run it from ModelScope first. I couldn't run it. I tried to run it from Hugging Face, wasn't possible. So, I went to their own Ling Studios studio, which is hosted, I think, in China. I tried to run a prompt here. So, I clicked on login, and then it just asked me for a Chinese phone number. Now, I don't have that, obviously. So, I couldn't really give that number here, so I couldn't run it in Ling Studio. I thought, "Okay, maybe I will just try it with API key. " I again tried it with my own phone number, uh the Australian one, didn't work. So, I had to have the Chinese phone number. So, they have made sure no one outside of China could run it, even with API. Anyway, so I reached out to someone in China, uh got an API key, and then I have uh run the code, uh sorry, run the prompt and obtain the code, which I will show you very shortly. And the whole code, I will also run it in front of you, so that you could see what this model has actually produced. But before that, I really want to vent a little bit about that, why only I mean, why even drop this monster at all if no one can run it, okay? You wanted to release a model in 1 trillion, fine. But why no small practical models like Gwen team, 0. 8 billion, 2 billion, 4 billion, or 8 billion, the way Gwen does? Gwen releases usable sizes that run on a single GPU and build massive community adoption. This feels like a pure prestige play. Who can who can really actually run a 1 trillion parameter model even with mixture of experts and only 63 billion active parameters, you still need a serious multi node CPU cluster that most developers and even mid-sized companies cannot afford. You know that I love to run models locally on the system and I just go to great lengths to run them having my own clusters even you know two node GPU but you know there is a limit which I can afford. But you know I couldn't run this model at all and why would you know they put it behind you know the you know Chinese phone number that is just absurd. Also another point is why would big corporations run this when they already have access to better closed source models like GPT 5. 5 Cloud Opus Gemini Grok and there are so many. They are cheaper in terms of API more reliable and require zero infrastructure hassle by the way. Even if you're in China you want to use a Chinese model there are great options in China starting from Queen's own larger models Deep Seek is

### [5:00](https://www.youtube.com/watch?v=Wg7Ln5tTsOA&t=300s) Segment 2 (05:00 - 10:00)

there ByteDance has done wonderfully well Tencent has models Baidu has great model Ernie. Why would someone go and run this RingLingMing models which are just coming in 1 trillion sizes anyway. So this was a Vengine under the hood it's a mixture of expert architecture with hybrid attention for long context. I will be talking more around its architecture but for now let's try to run it. So this is the code which it has generated let me actually show you my prompt to which I you know sent across to the model. Okay so if you look through this response I think this is worthy of a 1 trillion parameter model. So, I asked the model to act as an expert aerospace embedded software engineer and create a complete production grade ready to run flight control system for a simulated quadcopter. It had to be full project that can be installed on my fresh Ubuntu 22. 04 with one setup file. Including 6 DOF physics simulation, realistic sensor simulation like IMU, GPS, barrel estimator, cascaded PID controllers, flight modes like stabilize, altitude hold, lot of things. Mavlink support. I also gave it the you know, the whole hierarchy, folder structure, and then it should have some logging, some arming logic, and a web dashboard. So, this is what I gave it to the model, and it has created all these files on the left-hand side as you can see. Produced a read me, and then also the setup file. And these are all the source code which I received. All the components are there on the paper. This thing I have to give it to the model, and then let me go to my terminal where I'm running this Ubuntu system. Let's build this. And I'm just setting the permissions. It is installing the dependencies. So, let's wait for it to come back, and then we will go from there. So far so good. I will let it run in front of you. And now it is running, which is very good. I will let this uh screen open. I will go here, and I will run the dashboard. Dashboard is also running, no errors. And let's open it in the browser. I will reload this. There you go. So, it seems what it has done here, it has created the whole code. It has even created the web page, but it had just created a sort of placeholder. I have went through the code. In the source code, it seems that, for example, I'll just go with the sensors. Nothing in there. Mixer, nothing, nothing, nothing. So, this is all placeholder. And if you go here, file they're all there. You see, it has created some files. So, a lot of in some code it has created but missed a lot of it. And by the way, I am told it took around 30 to 40 minute because it was uh run with deep think. I'll actually quickly show you here. So, you go here in the link studio. There are different modes. So, click you click here. There is a deep think, there's a tool called So, it that uh prompt was ran with deep thinking. I don't know how much the cost was. I will check it later, but you can see that code is incomplete and it is it lacks real dynamics, um proper controllers, you know, no estimator, no mavlink, no advanced feature is there. So, seems like a working So, a temporary skeleton to get something running quickly to display instead of a proper full production grade flight control system that I asked for. So, this is a 1 trillion model which, after much ado, I was able to, you know, get access to indirectly and still didn't get it done properly. Anyway, so now you know the pains which you know, I had to go through for this. Coming back to the architecture, the focus seems very heavily on these coding capabilities with some stable tool calling, error correction, and multi-turn execution. Didn't happen in this case, as you saw. And as I showed you on the benchmarks, you know, high mode scores 87. 6 on pinch bench, beating some closed models. And X high does well on math and reasoning test. They also have built some new asynchronous RL training system with their Ice Pop algorithm to make trillion scale reinforcement learning more stable. Impressive on the paper, but you know, you saw just saw in action that really uh just seems like a technical flex showing

### [10:00](https://www.youtube.com/watch?v=Wg7Ln5tTsOA&t=600s) Segment 3 (10:00 - 10:00)

this lab can push open source to frontier scale. It might be interesting for researchers or enterprises like this inclusion AI with very huge budgets to spare. Or I don't know, seriously. Maybe you know, we would be able to access it through open router. But I don't really want to spend uh more API calls and time on this one. And if I have to spend API credits on a model, I would spend it in Claude or DeepSeek, or you know, some uh ChatGPT type models, not on the ring models. Let me know your thoughts in the comments. Please follow me on X for AI updates. And this is the channel if you'd like to subscribe. Thank you for all the support.

---
*Источник: https://ekstraktznaniy.ru/video/50840*