Coding with Minimax M2.7 in Claude Code

追加: 2026-05-11

[00:00:00]Hi everybody, Julian here. Today I'm trying out Minax M27, an openweight model from China as my coding model inside Cloud Code. I'll show you the setup. I'll run a full code review on one of my open source projects and I guess we'll send a pull request in the end. Let's get started. If you want to know more about Miniax M27, I recommend reading the blog post from Minimax. And as usual, I will put all the links in the video description. But because the model is openweight, we can also find a lot of good information on the hugging face page. Let's take a look. M27 is open weight and it's a mixture of experts model with 256 experts and eight active experts per token. That's about 228 billion parameters and about 10 billion active parameters per token.

[00:00:57]Weights are natively FP8. Uh these are the weights released by Minax. This is not a post training quant. Context size is about 200K. Uh the license is a custom Miniax license uh which you can of course read and should read before building with the model. In terms of benchmarks, again you'll find numbers on the hugging face page and in the blog post. These numbers are self-reported.

[00:01:24]For example, uh on SUB Bench Pro, the Miniax M27 model scores 56.22% which is within about a point of GPT3 codeex and uh on Vipro it scores 55.6 which is almost identical to Opus 46. So we see the model is definitely up there with the best uh closed models and well we'll see how it performs on my codebase. In this demo I'm using the plus highspeed plan. Uh plus highspeed is a paper token plan that costs $40 a month or $400 a year. So here's what you actually get on plus highspeed. You get 4,500 coding requests every five hours.

[00:02:16]routed through M27 highspeed which is a faster variant of the model uh faster than the standard tier and on top of that uh you get 9,000 characters a day for speech and 100 images a day for image generation. So if you want to play with uh the multimodel uh stack in Minimax uh you don't have to pay extra for it. It's included in the plan. So, the good news is um because you're watching this video, uh you get 12% off any token plan tier. Just go to the link in the video description. Isn't that great? So, how do we set up Miniax in cloud code? Well, this is really simple.

[00:02:58]Miniax exposes anthropic messages compatible endpoint, which is a drop in replacement for the anthropic endpoint.

[00:03:06]So we basically just point uh cloud code to a different URL and with a different key. So when you sign up for your plan, you can generate an API key and all you have to do is just edit your cloud code settings, the settings.json file, uh and use those exact lines here. And of course, just replace the key with your exact key. And that's all there is to it. Um I have to say during my session um it worked flawlessly. I didn't see a single error uh using um the minimax API with cloud code and and my plugins etc etc. So yeah very simple and as simple as that. Um, in this demo, I'm actually going to run everything in a Docker sandbox, uh, using the SBX command. Uh, and that lets me give the agent full permissions without having to worry for my local machine. All right, let's get some work done. Uh the repo I'm running this on is an open source project that I built called Cache Commander and it's a Rust uh terminal UI that scans developer cache directories, pip, npm, homebrew, hugging face caches and it looks for um known CDEs in your packages and of course it helps you find and easily delete all those unneeded packages and reclaim gigabytes or tens of gig GB of disk space. I use Claude uh to build this uh Opus 46, Opus 47. So, it's going to be interesting to run um M27 on this codebase and run a code review and see if it finds something on top of an already reviewed codebase. Okay, so let's uh start this session. On the left, I'm going to run uh Claude setup with Opus.

[00:05:05]And on the right, I'm going to run cloud code setup with minimax. And that's why I'm passing a different cloud directory here because I have those uh settings uh JSON lines that you saw just a second ago. Okay. So, we'll just make sure here we're using office 47.

[00:05:29]And what about this one?

[00:05:33]Okay, so this one we see minimax is selected. So let's use the two models here. Um let's maybe plan the code review with Opus and let's uh execute it with Minimax, which is actually one of the recommended ways to use this model.

[00:05:50]Okay, so I'm going to go to plan mode and say run a full code and test review on this repository.

[00:06:03]So, Opus is going to do its thing. It's going to plan the code review, maybe launch some agents, and then we'll get a plan in the end, and we'll use the plan with Minimax. All right, so Opus is done. It did a fairly extensive review of 22 uh cache providers in the tool, so there's quite a lot to to look at. I asked for findings only, no remediation techniques. Um, we'll let Minimax figure it out. Okay, so uh we have the path to that plan and now we're just going to pass it to Miniax and uh let it actually fix. Okay, so now we can just take the plan that Opus wrote and say based on the findings fix issues in the repo and dispatch agents to parallelize work. So it's going to read the plan. It's going to go through all the findings and uh pretty sure it's going to fire up agents to fix um all the issues that it found. So, it's actually launching a lot of different agents to fix those different problems.

[00:07:10]O2 tree performance. So, that's bad.

[00:07:13]Fix URL encoding, etc., etc. So, now we're doing all those things in parallel and my codebase is getting better. All right, we're done. 30 minutes later, we have fixed all the issues that were part of the plan. We've run all the tests.

[00:07:34]Everything is clean. So, why don't we create a pull request and send it to the repo. The pull request is ready. Number 33. Let's take a quick look here.

[00:07:48]Yep. See it here.

[00:07:52]And it's got all the right details, the bug fixes, the tests.

[00:08:01]Okay, now we have GitHub CI running.

[00:08:05]Yeah, looks good. And well, I guess I'll merge it and maybe release a new version based on this. And maybe to close things here, let's look at the usage this generated. until I started at zero and well I did consume 298 queries out of uh 4,500 for 5 hours so I should be perfectly okay. Uh and uh yeah, it's unlikely doing this kind of work you will ever hit usage limits. Um and uh remember that's only $40 a month. So this looks like a very good deal. So there you go. uh Miniax M27 in cloud code. Uh I really like the fact that uh the compatibility looks excellent. Um it was very easy to configure cloud to use minimax. I didn't see any tool error or any kind of error. Uh just works nicely.

[00:09:02]Um as you saw, we can work with uh maybe opus on the side and then um M27.

[00:09:09]Um very easy to do. use the best models for the best tasks. Speed felt very good. Um, I would say quite faster than than Opus. I didn't measure it, but just my gut feeling was that uh M27 was uh was a bit faster. So, that's it for today. Um, thank you for watching. If you want to try M27 on uh the plus highspeed plan uh or any other token plan for that matter, remember the link in the description will give you 12% off. So that's pretty cool. And all the other links, the model card, uh well the cash commander repo and so on. All of this will be also linked in the video description. Thank you very much for watching. I hope this was informative and fun. And until next time, my friends, as always, you know what to do.