MiniCPM-V 4.6: The FREE Tiny Vision Model Your Agents Need | UNBELIEVABLY REAL!

Added: 2026-05-20

[00:00:00]What if I told you there's a new AI model that can look at images, read text, and even understand videos, and it runs directly on your laptop. No cloud, no subscription, and nothing fancy. It's called Mini CPM V4.6, and it just dropped a couple of days ago, and it's tiny. It's only about 1 billion parameters. But here's the wild part. It actually beats Quen 3.5.8 billion model on more test. And it almost matches the performance of a 2 billion version, which is actually twice of its size. And the interesting thing about mini CPM v4.6 is that it's really fast like really fast. That is the time it takes to respond or let's say give you the first token is just 75 milliseconds. And when compared with similar Quen models, it takes almost twice as time. And this video we're going to take a look at mini CPM V4.6.

[00:00:47]We'll go through all the benchmark results and also why it is special and go through all the highlighted features of this model. And not just that, I'll show you how you can actually run the model directly on your computer and even on your smartphone. Okay? And as this model is actually very tiny but very capable and efficient one, you can actually run it on your computer in the cloud or even on your phone and it's actually really fast. And actually I tried using this model directly on my phone and this model is crazy fast. I uploaded an image and I asked a couple of questions about the image without internet or anything as such and the model was extremely fast. And if you're looking for like a model to process hundreds or thousands of images or videos, well, this is the model that you can use directly within your computer.

[00:01:28]It's very efficient. It's very cheap and only requires very little computing resources. And the interesting thing is that this is a multimodel, you know, vision model, but it not only supports images, but also videos. And as you can see, it is the smallest model in the V family with only 1 billion parameters.

[00:01:43]And it has a 2.4 times more throughput when compared with Quen 3.5, you know, 0.8 billion version. And also the TTFT that is the time it takes to generate the first token is significantly lower.

[00:01:54]That is it only takes 75.7 milliseconds and when compared with the competition it's almost like 2.2 times faster. So the mini CPM V4.6 model is actually the latest one in the family and it is actually used by a lot of real world companies like BMW, Volkswagen, Samsung, Horner, Lenovo etc. And again smaller models means lower deployment barriers, faster inference and border compatibility accelerating multimodel AI across edge devices. So usually if you want to let's say use a heavy multimodel you know and capable a models then you'll have to host it on the cloud and then use it and the builds may surprise you but the interesting thing about using mini CPM is that it actually requires very less resources to run and you can even run it on your phone directly as you can see it's a smaller and edge friendly that is it is a 1 billion parameter variant model and again you can run it on laptops and phones and real ondevice deployment without the cloud without even in the internet you can locally run it on your computer or within your local network or whatever, however you want to use it.

[00:02:51]And the thing is that it actually outperforms the Quen 3.50.8 billion version right here with high accuracy across most visual language tasks and higher comparable runtime efficiency. And not just that, it is low latency on device plus high throughput on the cloud optimized for either deployment. So if you want to use it in the edge in the device itself, you have options for that. And if you want to use it on the cloud, you can do it that way.

[00:03:13]In either cases, you will get low latency and high throughput. And when directly compared with quent 3.5 0.8 billion variant as you can see from the chart on total token throughput and prefill throughput benchmarks as you can see mini CPM v4.6 actually tops the chart by quite a margin. And again TTFT that is the time it takes to generate the first token is just 75.7 milliseconds which is actually 2.2 times faster or let's say 3.6 times faster when compared with the 2 billion variant. And here I have opened up the hugging face page of mini CPM v4.6. 6 as well as the GitHub page. And if you scroll down and if you click on mini CPM V4.6, you'll be able to read and find all these important details about this model. And again, you can find a full benchmark results test in here. So mini CPM V4.6 is being compared to Quen 350.

[00:04:03]8 billion, the Gemma 4 and also LFM 2.5, VL 1.6 billion. And as you can see, if you compare the scores in here in pretty much all these benchmark test, Mini CPM V4.6 six outperforms all these models up in here. Okay. And again, if you want to take a detailed look at all these benchmark test, you can go ahead and read through all these content right here. And as I mentioned earlier, you can even run the mini CPM V4.6 model directly on your phone, that is Android, iOS, and Harmony OS, which I'll show you later in this video. And as for setting it up on your computer, you have a couple of different options. So, I followed this guide right here to set up mini CPM v4.6 via Olama directly on my Mac. And if I open let's say terminal and I can run this dot/ola serve and I can hit enter. And right now okay it is running. Next step I can just go ahead and start the open BNB that is mini CPMv4.6. I can hit enter. And I just want to show you how fast this model is.

[00:04:58]So I can simply go ahead and let's say put a message like hi in here and hit enter. And there you go. It says of course I can help you with that. What would you like to know about the topic?

[00:05:08]Okay. Next step, let's just go ahead and try to let's say give a image to this model and see how fast it can respond.

[00:05:14]For example, let's just say this right here is the photo that I want to give to the AI model and I want the AI to actually give me all this text in a copyable form. Okay, and this right here is actually a image. So I can simply go ahead and let's say copy the image and then I can paste it in here and hit enter and as you can see it says added image and there you go. It is actually extracting all the textual content and see how fast it was. And again, mind you, this is actually a model that is being running locally on my computer.

[00:05:42]Even if I disconnect the internet, this model will still work. And right now, the model is actually giving us a summary of the content of this particular page. Next up, I'll go ahead and say, can you please extract all the text from the image and give it in here? And I can hit enter. And watch this. There you go.

[00:06:01]It is actually I mean like how fast it is right even though it is a model that is locally running with my computer. See how fast it is and the accuracy is actually topnotch as I can see it is actually very accurate and even it has got all these punctuations correct as well. Okay. And next step we can go ahead and try to give a photo something like this. This is actually a handwritten note. So I'll go ahead and copy the same paste it in here and I can hit enter and let's see if it is able to get it correct. Okay, let's just compare the output that is hello. Okay, simply noted has developed incredible proprietary robotic technology to write your message and envelope with a genuine real pen. See, so it doesn't matter if it is like a handwritten note or a printed one, the model is able to get it right. And also the speed is amazingly fast. And again if you do want to set it up all you got to do is to head over to this page right here and you can set up Olama then clone this GitHub repo check out to this branch right here and you can just go ahead and run this particular command that is amama serve and then you can run the mini CPM v4.6 or even the thinking variant by giving this command right here and not just that if you go to the hugging face page of this uh model you have an option for running it directly within the spaces. So as you can see here we have the mini CPM v4.6 demo and if you click on the same it will open up a page where you can run the demo. So I can click on this option right here that is the launch demo and after this I'll also show you how you can actually run it directly on your phone and use it without the internet as well. So wait for it. So let's just wait for it to load first. So there you go. The model is now loaded and here we have the option for adjusting all the parameters like max new tokens, temperature, top t all of that can be uh let's say adjusted in here. Next up I'll try to let's say attach an image. For example, let's just say this one right here. And now I can go ahead and say extract all the text from this image and I can hit enter. And there you go. It is really really fast.

[00:07:57]So that is one thing that I like a lot.

[00:07:59]And there you go. Here we have got the extracted content. And yet another interesting thing is that this model not only supports images but also videos.

[00:08:06]And I just went ahead and uploaded this video right here. So this was the video that I just uploaded as you can see like a sample video of a street in which a lot of people are walking around. And now I uploaded this uh video and I just told and gave a prompt asking to describe this video. And as you can see it says the video showcases a bustling urban street scene centered around a prominent church with a tall pointed steeple that stands out against the sky.

[00:08:30]And as you can see here we have the description. So basically you can upload both images and videos and use the mini CPM V4.6 model. And the thing is that it is really fast and you can actually run it on pretty much all the devices that you might own as of DAO. And next up, I'll actually show you how you can set it up and start using on your mobile phone. For example, in my case, I have an iPhone. So, you can just click this link right here. And again, I'll make sure to leave this link in the description below. And first of all, you'll have to get the test flight app.

[00:08:55]So, install it on your iPhone. And after that, you can just click on this link right here that is view in test flight and install the mini CPM V demo uh let's say app onto your phone. And next up, I'll move over to the screen recording part. So, there you go. Here I have installed the mini CPM V demo beta app.

[00:09:10]So, I'll go ahead and open it up. And when you open it up for the first time, the first thing that you got to do is to set up the model that is locally download the model to your phone. And for that, you can click on the settings button towards the top. And as you can see, Mini CPM V4.6 is already downloaded. So in your case, when you download the app for the first time, this model won't be loaded onto your phone. So all you got to do is to click on the model and then you can click on this blue button towards the bottom to download and load the model to your phone. And now that the model is loaded onto the phone, I'll actually show you how to use it. For example, maybe I can let's say attach u a image for example this one right here. This is actually again a handwritten note. So I'll go ahead and upload the same and wait for it. Okay. Next up I'll see what do you see in this image. Okay. And now I can hit enter. And there you go. As you can see it was crazy fast.

[00:10:03]So it has actually went ahead and extracted all the text from the image.

[00:10:06]And here we have the full output. Next step, I'll select this photo right here and click on finish. And I can go ahead and say describe this image. Oops.

[00:10:17]Image. And hit send. And let's see. So this image shows a digital screenshot of a fitness or step tracking. And again, we have all the highlights or that is the highlighted details in here from the screenshot. So this right is the screenshot that I attached. And this is the response that we got. And it is actually crazy fast. Even though this model is running locally within my phone, as you saw, it was like crazy fast. And again, the time it takes to generate the first token is just 75 milliseconds. And if you do want to run this model locally on your phone, you can directly head over to this link right here. Or if you're using, let's say, an Android device, I'll make sure to leave links in the description below where you can actually set it up directly on your Android device as well.

[00:10:55]So that's pretty much all I wanted to show you in today's video. So the mini CPM V4.6 6 actually is a very capable multimodel LLM that can actually see that it's basically like a vision model that takes both images and videos as input and it is crazy fast and the interesting thing is that you can directly run this model on your machine like phone or laptop without any high-end GPU or graphics card or anything as such and again I'll make sure to leave all the links to the all the resources in the description below so you can go check them out and yeah I hope you guys found this video useful if yes make sure to subscribe and I'll see you in the next