Mistrol is so back. They have just dropped this amazing Apache 2 licensed coding model which is suitable and geared towards agentic coding tasks. This is Fad Midza and I welcome you to the channel. In this video, we are going to install this Devstrol on our local system and we will see how it performs. The performance of this model also makes it a suitable choice for agentic coding on privacy sensitive repositories in enterprises especially ones subject to stringent security and compliance requirements and especially if you are in Europe and you have to comply with all those local laws Devstrol also makes sure that you comply with them. This model not only comes in Apache 2 licens syncing but also it seems that from the benchmarks which Mistl has released the performance looks quite good. You can not only obtain it from Mistl's own website but also from hugging face and that is where I'm going to download it. This model is a fine-tuned version of Mistl small 3.1. So it has a long context window up to 128k tokens as a coding agent devstrol is text only and before finetuning from mistrol small 3.1 the vision encoder was removed and we already have covered that mistrol small on the channel. So let's get started. I'm going to use this Ubuntu system and my GPU card is Nvidia H100 with 80 GB of VRAM. But we will see how much VRAM it consumes because as per ML you can install it on 40 GB of VMA but we will check it out for VM and GPU. I'm very grateful to Master Compute for sponsoring it. If you're looking to rent a GPU or VM on very very affordable price, you can find the link to their website in video's description. Plus, I'm also going to give you a discount coupon code of 50% for range of GPUs. Now, the tool which I'm going to use in order to get this installed is VLM. So, make sure that you have it installed. Let me not sorry this one. Let me quickly show you the command to check it out. So this is a version of CUDA and m you know vlm with mstall common I am using now all you need to do after you install torch and stuff just install VLM or just install VLM upgrade it is going to install everything for you so I already have it I'm not going to run it more if you want to learn more about VLM just go to my channel search it and I have done easy to understand tutorials on VLLM from every angle model. Okay, so our V llm is all good. Now, next we need to download and serve the model with VLM. And this is a command where we are serving this devstral model with VLM. We are specifying tokenizer. And then as we already got that mstrol common library which we are specifying in here in this parser. So let's run this. And the model is now being downloaded. As you can see, it is doing uh starting to download all the files. So, let's wait. And while that downloads it, let me also introduce you to the sponsors of the video who are Matrix. Matrix enables you to create a realistic world simulator where you can repeated uh where you can run repeated simulations on AB tests for your marketing captions allowing the marketers to test different angles and strategies and I will drop their link in video's description. And first time you just have to be a bit patient um patient because it takes bit of a time and you can see that it is downloading a consolidated file of over 47 gig. So I'm not entirely sure how that fits into 40 GB of VRAM as mentioned on their hugging face card. But anyway, we will check it out. Model is almost downloaded and model is now being served and the model is all ready and good in our V LLM as you can see. Let me quickly show you the VRAM consumption too. I'm very curious about that. There you go. So, let me run it in front of you. So, over 74 gig of VM. I don't think so it is going to run in this precision on your local system anything less anything uh I would say uh less than 80 GP of V RAM okay anyway doesn't matter so we already have everything let's do the inference for inference there are uh two things you need to do first up you would need to access uh their hugging face repo and there you will find a file called as system prompt just copy paste and use it here. So it just tells it what exactly it is intended to do. It's a helpful agenting model and there is a lot of other things like open hands scuff and stuff which is another model and framework which we already have covered on the channel. Now this is a file app.py where I am doing the inference. So I'm just importing few libraries. I'm giving it the path where it is running locally. So this is the default path of vlm. If you're running it on any other um endpoint just replace it. This is a model name. We are loading system prompt and this is a file which already have been placed here. And this is a prompt which I'm asking to do. So in this one I'm asking it to do to please generate a complete self-contained web page that displays the following a bouncing ball animation. So it's a very common animation which we normally do in order to test out the model. This is my chat template and this is my data and all that stuff with some hyperparameters. It is going to print out the whole code here. So let's go back to our terminal and then let me I'll just minimize it. I'll go here and I'm just going to run this code from here. Sorry, I'm just adjusting it. Let me clear it. And then I'm just going to run my app.py from the root of that folder. And there you go. It has given us the code. And you can see it looks quite okay to me. But let's check it out. And I'm just going to maybe quickly create a very quick index dot HTML here. We can paste it. Oops. Yep. this one. Let me copy and paste the code. I think it I didn't copy it properly from the terminal. And there you go. When I have opened it in the browser, this is what we have got which looks quite good to me. I'll just reload it. There you go. It's a bouncing ball and then it settles down pretty good. basic but but it it has done what we asked to do. Now the best use case of this model in my opinion is when you integrate it with tools like cursor continue and there are heaps of them now and I will be making further videos on those tools too but for now I just want to test out the coding capabilities of this model uh in this video. So in the next one I'm asking it to do a quite a tough question where it needs to implement a map reduce framework in pure python no external libraries or stuff and then I have given it some of the characteristics I want around clean code structure some of the documentation and reproducible text and example. So let's go here and then I'm going to run this. So let's wait to see what it produces. And there you go. It has given us a very good code I would say because it has just used the Python primitives and it has defined the classes in a proper format. The commands are there. It is properly instrumentated. The master class is there with all the decorators and stuff. Yep. Looks quite good to me. It has even implemented that map and reduce quite nicely. And then there is example usage. So it has followed it very very nicely. I believe this could be very well suited for your own MCP custom MCP uh servers where you would have some of your own legacy or custom application. And we can also test that code out by on a large file to see if map reduce works. Wow, very nice. Works pretty nicely. It has done the map reduction. Very basic, but still it did what we asked it to do. And finally, let's test out this um TypeScript uh for creating a new plug-in where I'm asking it to design me a hot reloadable plug-in system for a long running NodeJS application. And then there are some conditions which this plug-in needs to follow. Also, I'm asking it to only use nodes standard module and TypeScript typing and no external plug-in management frameworks. So, let's run this. Let me run this. So, wow. So, when it works, it works like a charm. Um, very very fast. So, if I quickly go through the code looks sublime to me. All the configuration is there. All the sockets are there. Even it is defining the pickles. All the functions, worker processes, asynchronous, all the ingredients are very very well aligned in this code. Good stuff. And if you are a TypeScript developer, please also uh let us know. And then there is example usage too. So look, stellar effort by Mistl I would say. Very impressed by the model. And as I said, I'm also going to make other videos with a lot of um coding companions. So stay tuned. Let me know what you think about this. If you like the content, please consider subscribing to the channel. If you're already subscribed, please share it among your network as it helps. Thank you for watching.