How to Use MiniMax M3 in Claude Code (Full Setup + Tests) | Open Source AI

[00:00:00]So, some of the most interesting things, at least in my opinion, in the world of AI right now aren't coming from OpenAI or Anthropic. They're coming from the world of open-source models, and one of those open-source models just hit 59% on the SWE Bench Pro benchmark. That is ahead of GPT 5.5 at coding. It has a million token context window. It can see images and videos natively. And here's the kicker, it costs about 15 times less than Claude Opus. And that's than Claude Opus. Forget the more modern right your Fable era models. It is significantly cheaper for that degree of performance.

[00:00:39]Now, that model is called Minimax M3, and in this video, we're going to plug it into Claude Code. Claude Code is the agent that I use every day. It's what you probably use. And so, we're going to put Minimax M3 inside of Claude Code through three tests: a real build, an entire codebase, and a screenshots-to-code challenge. by the end, you'll be able to understand and judge the quality of these open-source models. Do they meet the hype, and can they actually replace the closed-source frontier models?

[00:01:10]That's the big question of AI right now, and we're going to tackle a little part of that in this video. It's going to be a fun one. Let's go. Now, just quick context first on open-source and why releases like M3 matter. Basically, until this class of models, if you wanted native multi-modality, million tokens of context, and frontier level coding, you had zero options. The frontier closed-source models worked just better across the board. That would be fine, except for price, where the margins that the closed-source companies have are through the roof. And so, yes, the models are amazing. They're better than everyone else's, but the price reflects that. Claude Opus is $5 per million input tokens and $25 per million output tokens. So, if you're running especially agentic coding workflows all day where the model is reading through an entire repo, it's rebuilding a code base, that gets really expensive really fast, hence the term token maxing that's come out in recent months. And so, the open-source question that has started to arise is if I can get 95% of the performance for 5%, 1%, 0.1% the cost, is that not a better deal for me? And so, over time, the open-source models just kept getting better over the past year. DeepSeek, Kimmy, the MiniMax series, but they were still kind of iffy back and forth. They had really great context, but sucked at coding. They were amazing at coding, but missing something else. MiniMax M3 as an open-source model checks all of those boxes. That's how the AI community has reacted to it, and so that's why we're looking at it in this video. So, on that note, what is MiniMax M3? And then after this, we'll jump right into my computer for the demos. So, as the name implies, MiniMax M3 came after the M2 series out of the AI lab MiniMax. Again, it ranks 59% on the SWE bench pro. That puts it ahead of GPT 5.5 and slightly behind Opus at real-world engineering tasks. And then 66% on terminal bench 2.1. And then on browse comp, which tests autonomous research and tasks, it scores 83.5, which actually beats Claude Opus. Now, the context window is up to a million tokens of context, and that's with a guaranteed minimum of 512,000.

[00:03:30]And that works because of a new architecture out of the lab called MiniMax sparse attention or MSA. And basically, it just means your inference is cheaper, and so costs are going to drop. And then it is natively multimodal, so it can directly read things like images and videos, your visuals across the board because it was trained on things like that. So, it can do all of that as part of coding tasks.

[00:03:53]And then now, again to stress, this isn't open-source model. The weights are completely open. It is all on Hugging Face right now. It's about 428 billion total parameters with about 23 billion active per token. So, you can literally download the thing. You do not have to pay the companies to use this. You can go to Hugging Face and you can download it. So, now we're going to get into my computer. Now, this video thankfully was sponsored by MiniMax. Everything I'm about to show you is real. It is completely uncut. So, thanks for the MiniMax team for letting me showcase the model. And now, I recommend that you try this out for yourself. While you're watching the video, just follow along with what I'm doing cuz that means you'll learn better. It's a better use of your time. So, you can do that at platform.minimax.io.

[00:04:37]That is on the screen here or again platform.minimax.io in the description. Now, if you want to use the model through the official MiniMax platform, you can get 12% off using the link in the description. Or again, it's open-source and you can download it. So, if you want easier UI/UX, a better interface, you can use it through the platform and you can pay a monthly subscription or you can again use it in its full open-source sense.

[00:05:03]So, now let's go ahead and take a look at the model and wire it into Claude Code. So, that's coming up right now.

[00:05:08]Okay, so we're going to start off here with the install of the M3 model inside of Claude Code. And to do this, we'll be following this article. We'll put the link in the description if you want to see that and we'll go to terminal. Now, inside of terminal, we want to start by clearing the prior API keys for Claude.

[00:05:25]And as a whole high-level process, what we're doing here is keeping everything about Claude Code the same but switching the brain from the Anthropic models to the M3 model in this platform. So, let's use this command. So, unsets these two tokens. I've clicked enter. We'll run this as a quick check. Nothing came back. Looks good. And now, we're going to have to blur this but basically what we're doing is pasting this command right here to install the M3 model.

[00:05:53]Let's click enter. Now we can verify, that looks right. So that is the entire install right there. Again, we're basically just swapping out the keys. So now to test whether this works, let's just go ahead and type Claude to open the Claude 2 code portal. And then we can do forward slash model to see the model that's being used. And then here we can go ahead and see the default model MiniMax M3 down through the custom model right here. So let's stick with default and that will be the custom MiniMax M3 model. So it works and that in mind we can get into our first proper demo. So again, this looks and feels like Claude, but the brain behind it, despite the interface you're used to, is going to be the cheaper M3 model. And so let's start off by giving it this prompt, right? Build a full stack expense tracker web app, React front end with a dashboard, Express back end, and so on. We're going to host that locally so we can go ahead and see, "Hey, how good is M3 actually at programming?"

[00:06:55]Let's give it that prompt. Okay, let's proceed. We're going to work out of downloads. Okay, we are proceeding here.

[00:07:01]And if you want to check your usage, you can go to the MiniMax platform dashboard right here. Okay, it's going through these steps. It's creating the project structure, then working on the back end, and then finally the front end. Okay, there we go. It's all finished. We can see it's been saved in the expense tracker project. And then to open the app, we'll just paste this URL. So we're in incognito tab. Let's paste that. And here we go. Here is our expense tracker.

[00:07:28]Let's see if import CSV works. It does.

[00:07:31]And then we can also add transactions manually and store the transactions. So that gives you an idea of what it can do in terms of programming ability. But again, one of the big differentiators of this model is the fact But again, one of the big differentiators of this model is its context window. And so, for that, what I want to do is go to a repo. This will be the Fast API repo. Basically, it's very large. And so, what we're going to do is we're going to clone this. We're going to download it and then ask the model, "Hey, can you do some inferential tasks where you have to go to different parts of this very large database and understand it and interpret things about it and then see how well it can do that. So, let's head back into terminal. We'll go into a new tab. I'll paste this command to download that repo. That's done. Now, we're going to launch Claude inside of here and then we're going to give it this prompt. So, read the entire codebase and produce an architecture overview explaining how the modules inside of it connect and do some other work. So, let's give it that prompt and we'll let it read for a minute. And keep in mind, there's no rag workarounds, no chunking going on here.

[00:08:42]We just copied in the repo and we're letting it run. And we can see here it's set up It's setting up a task tracking and then reading those affected files.

[00:08:52]Okay, the refactor is completely complete. All 24 affected tests passed.

[00:08:58]We can see the architecture overview here, the core app and routing, dependency injection. The list goes on here with all the refactoring and the changes that were made. And then you can see right here we went ahead and committed that. So, it's done. It is fully good to go. Pretty impressive checkmark on the context window. And now, let's jump into the third demo we want to stress test here. And now, the last big point that we want to stress test on the model is the fact that again, it's natively multimodal, meaning it can understand images and videos. So, one useful flow I like with that kind of feature is rebuilding visual things that you like to see on the internet. So, here's a website of a company called Crusoe. So, I've opened up a new terminal tab here. Let's go ahead and get into Claude. And then what I'm going to do is screenshot this tab. I'm going to jump back into terminal. I'm going to paste the image and then I'm going to give it this prompt. Here's a screenshot of a landing page design. Rebuild it in this way. And so, we're directly feeding that image to the AI so we can understand and extract the information we want it to and we don't have to sit here typing out use this kind of font.

[00:10:12]This is the design for the page and so on. So, let's let it run. Okay, it went ahead and finished the image. Let's give that a test here. And there we go. This is a real page that was programmed by the bot. This is the real website. This is one prompt on M3 to rebuild it. And we can see it even built out the full page here. But this is really, really good, right? I mean, you don't I don't have to say it. This looks borderline exactly the same. What we don't see here are the animations and the movement, so that's what we would want it in a second follow-on prompt to add. But for one prompt, honestly, this is the most impressive thing out of what we've looked at and that I've seen out of M3.

[00:10:54]So, that's looking really good and that wraps up our dive into MiniMax M3. So, we've looked at performance. Now, the other side of that, of course, is going to be cost. And that like I've been saying with the whole open source trend is really where you start to see the argument for why people are liking open source over closed. So, it can perform at the caliber that we've taken a look at today. But here's the cost. 30 cents per million input tokens, $1.20 per million output tokens. That compares to $5 and $25 out from a model like Opus, much less Fable. So, that's a 15x price spread. 1X, 2X, 5X would be ridiculous.

[00:11:35]15X is next level and that is really again where if you're a CFO and it's like, "Hey, can I collapse my costs by 15X on half of my usage for a net 7X decrease?" That takes your bill from 10 million bucks a month to 1 and 1/2 million dollars a month. And now the way the pricing for M3 works is you can do it via API. You can also go off a subscription token plan. That's your 20 bucks, 50 bucks, 100 bucks a month options. And that's $20 a month for 1.7 billion tokens of M3 usage monthly. And then 5.1 billion for the $50 a month option. Now also the way that MiniMax works as a whole, this is a company with a lot of models out there. MiniMax M3 is the new one, but they also have 2.7, 2.5, and then your audio and your video models as well. All of that is wrapped into the same platform. So there you go.

[00:12:33]That's the model. I hope you judge for itself based on what you've seen in these demos. Again, those are uncut. We don't hide or remove anything. My take on it again is still closed-source models are better, right? Opus across the board is going to be better for tasks as well as Fable. If you need absolute maximum capacity and performance and you don't care about cost, those closed-source models are what you should go with. I think the fundamental argument of the MiniMax M3 model that you saw here and open source as a whole is, "Hey, if you can pay 5, 10, 15, 20, 50 times less, whatever the number is, for 90 plus percent of the performance to be there, is that a worthwhile trade for you on some or most tasks?" And for nearly all of you, for me, the answer is absolutely yes. And if you do intend on self-hosting the open weights, that does take some pretty serious hardware and you should check the license terms there for any kind of commercial use. That's just standard disclaimer for whatever you're doing in the open source world. So, the token plans in the API through the platform are going to be what a lot of people want to do. If you do want to go all the way in to open source, again, go to Hugging Face, download the thing. So, overall, I think this is for any of you and myself, to be honest, who spend a lot of money on tokens. You do it with all the frontier models, and there's that meme of like using fable or you're using a flamethrower to light a match.

[00:13:59]That's kind of what it is to use the leading closed source models on all of your tasks with AI that you let agents do for you. So, that's where you're at.

[00:14:08]Then, just start to cycle in models like these. You'll save money, performance will be the same, life will be good. And so, that in mind, again, that's the video. I hope you understand this model now, and you can go, again, to platform.minimax.io to support the channel and to give it a go yourself. And you do get a 12% off any token plan if you do that. So, drop a comment with what AI model you want us to take a look at next. That in mind, I appreciate you all for being here, and I will see you in the next one. Cheers.