Xiaomi Launches MiMo-V2.6-Pro as Top Open AI Model: Let's Verify

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: Xiaomi Launches MiMo-V2.6-Pro as Top Open AI Model: Let’s Verify · Channel: Fahd Mirza · 13:17 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, AI models, Multimodal models

Summary

The video tests Xiaomi’s MiMo-V2.6-Pro, a large multimodal model trained with a combined reinforcement-learning approach, on coding, design, and vision tasks. It fixes a subtle passenger-counting bug in a ferry occupancy dashboard, though the run takes a long time and uses many tokens. It also produces a distinctive coffee-shop website and turns an image into a 3D game concept and prototype. The prototype’s bounce mechanic works, but its movement keys remain broken after the model attempts a fix, leaving the results mixed.

Key points

MiMo-V2.6-Pro is described as a 1.02-trillion-parameter mixture-of-experts model with 42 billion active parameters and a one-million-token context window.
The model combines coding, agentic, visual, and cybersecurity tasks in a single reinforcement-learning run.
In the ferry dashboard test, it finds and fixes a bug that counted bookings rather than passengers, correcting the displayed occupancy.
The dashboard fix succeeds, but the model takes more than 25 minutes and consumes many tokens.
The coffee-shop landing page has a distinctive editorial style, working navigation, and creative copy.
In a vision test, the model identifies objects in an image and maps them to interactive game mechanics.
It generates a playable 3D prototype with a working bounce mechanic.
The prototype’s WASD movement controls do not work, despite the model attempting to fix the issue.

Chapters

Start	Chapter	Summary
00:01	MiMo-V2.6-Pro and the ferry dashboard test	The presenter introduces the model and asks it to diagnose a passenger-counting bug in a live-style ferry occupancy app.
02:45	Benchmarks and reinforcement-learning background	The presenter discusses benchmark claims and the model’s combined reinforcement-learning training before the app fix completes.
05:29	Verifying the bug fix and testing web design	The dashboard now shows the correct passenger count, after which the presenter evaluates a generated coffee-shop landing page.
09:17	Image understanding and 3D game prototype	The model interprets an image as game-design elements and creates a playable 3D prototype from the concept.
11:31	Testing the prototype controls	The presenter reports that the bounce works but movement keys remain unresponsive, even after a requested fix and checks in other browsers.

Quotes

This model can handle text, image, video, and audio all in one model with a 1 million token context window for long agent runs. — 00:36

And now you see it says booked 92. — 05:29

this was a very subtle and nuanced bug which it has fixed. — 06:00

It identified every object in the image accurately — 10:21

Bounce is working, and I have hit it. — 12:05

Concepts and entities

MiMo-V2.6 Pro, Xiaomi, Hermes Agent, Nusantara Ferry Lines, reinforcement learning, Deep Sue

Full transcript (original language)

00:01 It seems everyone in China is building AI models from food delivery services to giants like Alibaba, Baidu, and various others. And now we have Xiaomi also chipping in for quite some time and we have been covering their Mimo and other family of model series for quite some time. And they’re not bad either. This time they have come up with V2.6 Pro in RL with various flavors. This is their flagship model 1.02 trillion parameters mixture of expert

00:36 with only 42 billion active. This model is built to scale reinforcement learning towards self-improvement. It mixes coding, agentic, visual, and cybersecurity tasks in one reinforcement learning run instead of training them separately. So, skills transfer across domains. This model can handle text, image, video, and audio all in one model with a 1 million token context window for long agent runs.

01:07 Today I’m going to point it at a broken production style app and handing it the wheel to fix the bug in this one. On your screen is Nusantara Ferry Lines, a live occupancy dashboard for inter-island ferries in Indonesia. Every gauge here is lying to you. For example, Merak to Bakauheni shows 5% full with 114 seats free, but that ferry actually has 92 passengers booked and only 28 seats left. The booking endpoint counts

01:42 booking rows, not passengers per booking. So, a single booking of 30 people counts as one. The dashboard massively underreports occupancy. In the real world, that’s an overbooking and safety risk. Staff would sell seats that exist. And you know what? If you have been keeping an eye on the news from Indonesia, there have been a lot of fairy incidents. Uh thankfully, they are taking care of it now, but still it’s a real-world issue where this application is actually

02:14 running locally with a back end with front end full stack app with lot of components as you can see here with a real-life database. What I’m going to do after integrating it with the Hermes agent, I’m going to give it this goal, which is not telling this model what exactly the bug is, only asking it to go and fix after finding it by itself. While it runs, let’s talk a bit more about the benchmarks.

02:45 So, you can see that as per this image, it is punching right in the same weight class as Claude Opus 5 and GPT-6 Astra. Beating both on real coding benchmarks like Deep Sue and topping the center field on job bench. It doesn’t win everywhere and Opus 5 is very old. But I think for a model this size going head-to-head with GPT-6 Astra is a real good deal. But rest of it doesn’t seem too impressive. Um but the scores are not bad, especially when you

03:18 compare it with Fable 5.1, you see that Fable 5.1 for example to let’s run very far. It is very close. Anyway, I really don’t pay much attention to these benchmarks. We are more interested in how exactly it performs. So, let’s wait for it to finish. By the way, recently I have started this free weekly AI newsletter, which you can subscribe at fahadmirza.substack.com. And if you want to help out the channel, please go to the home page and become a member. You will also find the link in

03:50 first pinned comment. And this is the same model which made quite a headline when they streamed it live. And not only that, what they did was um one mixed reinforcement learning run across coding agent vision and cybersecurity together as I said instead of separate runs per skill. They also scaled the reward signal itself and AI grader ranks passing solution against each other instead of just pass fail closing a self-improvement loop. And then they streamed this whole actual

04:24 production RL um live as it happened in under 6 days. Mimo V2.6 Pro completed 30 RL steps over roughly 750k trajectories for about uh 2.62 million US dollars lifting its Deep Sue 58.4 to 72.57, which is a 14-point jump that generalized well beyond the training task. It is almost there and as you can see there it has

04:57 um found a bug. That is fine, but I think it is wasting a lot of tokens. Taking long time. It’s already more than 25 minutes or so. It is still working through it. It is not that hard a bug. For instance, I just did Grok 4.7 um this morning and it was quite fast and didn’t waste too many tokens because at the end of the day these models are now plateauing. And this is where we need to see the

05:29 difference at how quickly and how cheaply they find and fix the bug. So, it has completed the goal. I would just let it come back and then we will test it out. Okay, goal completed. Let me hard refresh this page and we will check out. So, for instance, keep an eye on this booked figure here. It says booked six. It should jump. Let me control F5. And now you see it says booked 92. And this one has also jumped. And now

06:00 the actual seats are being shown which are available. And that should avoid any catastrophe. But look, I think the model is not bad at all cuz this was a very subtle and nuanced bug which it has fixed. Brought the stack up, but as I said earlier, took long time and just consumed a lot of tokens. Anyway, let’s do another test. Let’s do another test. So, if you look at the prompt, we are handing Mimo a detailed creative brief for a Melbourne coffee shop landing page, bold colors,

06:32 editorial layout, real personality, none of the generic template defaults most models fall back on. So, let’s see if it actually delivers a design with taste because that is what they say in their blog post that this model has got a taste for front end. So, let’s see if the taste is there or not. I already have given this prompt to the model and it is working through it. It’s going to take bit of a time. So, let’s wait. Meanwhile, as I said earlier, please subscribe to this newsletter which I release for free

07:03 weekly for hamza.substack.com. And if you want to help out the channel, please become a member. So, I will wait and then show you what it generates because I’m also asking it towards the end to start a local Python server and then confirm it’s running and give me a URL. Okay, so the model has come back, started the Python server, created that file, and this is what it looks like. It has even given me the port. So, you know what? I’m just going to run it as is.

07:37 I have just clicked on it. Or maybe I just need to copy it because this is a Ubuntu machine. So, I’ll just grab this and run it in front of you. And there you go. So, this is what it looks like. Let me try to minimize this. Okay, so this is a duck board. This is the name of the cafe, by the way. And menu, story, find us. You see, menu works. Story works, and find us also works. So, I’ll just go at the top and we’ll

08:13 scroll through. The font looks good, sort of a chic style. Uh do you think it’s a modern one? I’m not sure about that. And see what’s pouring, I’ll check it shortly. And then rest of it, there is a menu. This one looks good. This is pretty different. And then it is talking about its story. And this looks good. You see, it says that we got evicted

08:44 from perfectly good garage and needed somewhere to put 9 kilos of Ethiopian dream bean. Very nice. And this is nice. Quite modern. And then very nice. You see, this is the creativity that if you are coming to this place, if you’re coming through tram 86 to tram stop 20, and then vendor and 130-ish. So, this is quite good. Encouraged biscuits behind the counter. So, creativity is quite good. Don’t

09:17 doubt about that. This one is also not bad. I’ll just go up and we’ll click on that button which I wanted to earlier. So, this is the board. Very nice. So, not bad at all. Uh but I will also take your advice on it. Let’s do a vision test now. Hello, by the way, it also creates the skills by itself. You see, I didn’t ask it, but it created a skill out of it. Let’s do a very hard vision test. So, and this is right according to their

09:49 blog post which they claim. I’m testing if the model can comprehend this image and then, more importantly, if it can turn that into a structured game design logic, which is the actual claim. So, this is the image that I already have given to the model and this is a prompt to create me a small interactive 3D game scene concept based on it. And then it needs to describe the environment and that all that stuff. So, let’s see what sort of output it provides because I have given it free hand here.

10:21 And the model has come back with a really strong result. It identified every object in the image accurately, like balls, mice, post towers, donut bed, and lot of other things. And mapped each one to a concrete interactive mechanic instead of just describing the scene. So, as they claim that it can decompose a request into multiple tasks, that is true. Now, let me see if it can actually generate something out of this concept. So, I’m asking it to build me a simple playable 3D prototype in a single HTML file. Let’s see what it produces.

10:59 Okay, so the model says that it has completed the goal. Let me open this URL in browser in front of you and we will check it out. There you go. So, okay, now it says there is something. Oh, wow, this is not bad. So, it says that space to bounce. Yes, it is bouncing and this is not bad at all. Wow. This is so good. It is even jumping. And you can see the rest of it is quite

11:31 good. So, the bounce is working with the space bar, but the problem is that my keyboard keys w a s d are not working. By the way, w is forward, a is left, s is back, d is right. These are the keys, but when I press them, nothing happens. So, if I, for example, press F12 in the developer mode, you see, this is the error I’m getting. So, what I’m going to do, I’m just going to go back down here, and I’ll tell it this. That the game has this bug. This is the

12:05 error. That fix it. Let’s see if it is able to fix it or not live. Okay, it says that it has fixed it, so let’s me do hard refresh with control F5. And now we have the red ball. I’m going to move with keys. Still not nothing is happening. Bounce is working, and I have hit it. That is good. I’ll just keep moving. Okay, so keys are still not working.

12:43 Maybe it is my browser issue. I don’t know. Maybe I’ll just press F12 again. And I’ll focus. I don’t see any errors. This is not bad at all. You know what? Let me open it in another browser and try again. Just a sec. And now this is Microsoft Edge. I’m going to press space bar. This is good. No, the keys are not working. I also have checked it in Chrome. But keys don’t work. Bounce work. Other

13:17 than that, I think it looks okay. Not bad at all. In terms of image understanding and actually producing that simulation or 3D animation. So, that’s it. That is Xiaomi’s Mimo for you. There are two other models. There is also a distillation of coin, which I might cover uh in the next couple of days, so stay tuned. Let me know your thoughts in the comments. Thank you for all the support.