MiniMax M2.1 Is Coming! Open-Source Agent AI Gets Better

A few months ago, an open-source model quietly became one of the most interesting agent focused releases we've seen this year. That model was Miniaax M2. It didn't try to compete head-on with the biggest generalpurpose chat models. Instead, it focused on something much more specific and much more useful for developers, agents, tools, and real workflows. At the time, M2 didn't get attention because it was flashy. It got attention because it actually worked where a lot of models struggle, especially on long horizon tasks that involve planning, tool use, and execution. Now, Miniax is preparing a new release called M2. 1. In this video, I'll quickly recap what made the original Mini Mac 2 important. Then, I'll explain what's changing with the upcoming M2. 1 release, and I'll finally show you this model in LM Marina. So, let's get into it. To understand why Miniax 2 mattered

you have to understand what it was designed for. From day one, the team behind Miniax had a very clear philosophy. Intelligence should be usable by everyone, not locked behind expensive subscriptions and not optimized only for chat. M2 was built as an agent and code native model. That means it was optimized for tasks like writing code, calling tools, planning multi-step actions, and executing workflows that don't finish in a single prompt. This is why M2 performed especially well in agentic evaluations even if it wasn't always the strongest. Another important detail is that Miniax M2 was fully open- source including model weights. Developers could run it locally, deploy with VLLM or SG lang and integrated directly into their own systems. On top of that, the pricing was extremely aggressive. At launch, M2 was roughly only 8% of the cost of Cloud Sonnet while offering nearly double the inference speed in many setups. That combination alone made people pay attention. The Miniax team often talks about what they call the impossible triangle. In most AI models, you can only optimize for two out of three things: performance, price, and inference speed. High performance overseas models tend to be expensive and relatively slow. Cheaper domestic or open models often struggle to match performance or reliability. Miniax M2 was an attempt to break that trade-off. Instead of chasing maximum benchmark scores, they focused on building a model that was good enough, fast enough, and cheap enough to make agent-based systems practical. On aggregated benchmarks like artificial analysis, M2 ranked among top open models globally while still remaining dramatically cheaper to run. In other words, it landed in that rare green zone where performance, cost, and speed are all acceptable at the same time. Now, let's talk about Miniax M2. 1.

This is not a brand new architecture or a reset of the model. Instead, M2. 1 is a refinement of what already worked. The goal of M2. 1 is to smooth out quirks from the original release and expand usability for developers building more complex systems. One of the most important confirmed upgrades is multi-coding language support, which makes the model more practical for real production environments. In addition to that, Miniax has confirmed that the official API support and structured coding plans are launching very soon. This matters because it signals a shift from interesting open-source release to something closer to a stable developer platform. Iteration like this is usually a good sign. It suggests that the model is actively being used, tested, and improved rather than abandoned after launch. So, I'm on Elm Marina right now

and I'm going to test out the new Miniax M2. 1 preview model on here. So, what I'm going to ask the model to do is make me a Voxal Pagota garden where I can rotate the view, zoom in and out. And I want to see the model's capabilities on two things. Number one, its capabilities on following the instructions and number two, does the garden actually look good? So, let's test this out. All right, looks like our model is done creating the pagota garden and you can see that it thought for about 2 seconds and it created a plan. So it's going to first create a single index html file with 3. js. Build a vauosaw garden with pagota buildings using box geometries. Add controls for mouse touch interaction to rotate and zoom. Include garden elements like trees, water features, rocks. And then you can see that we actually have a pretty good looking garden. So we can see that we can move around. looks pretty good. The pagota looks like it's levitating a little, but not bad. Like it looks pretty nice. And we have like trees here. Everything like that. We also I really like this uh feature of like that. What do you What is this? Like I guess it's trying to be fish over here. Like it looks like there's a fish pond and everything like that. You can zoom in and out. Like I can zoom in. Let me zoom into the fish, I guess. Like let's see if I can zoom in specifically over there. Uh, looks like I can. And then we also have a feature that I added by itself, which is a toggle day and night. So, we can make it night. Okay. So, night is obviously very dark. All the features look pretty dark. And but there's a light. Oh, it looks like it added a light feature here, which is pretty cool. And then we can reset the view. And then let's make it night. So, this is pretty good. And it also has a legend at the bottom here. Red roof tiles, wooden structure, trees, and grass and water features. So, this is not bad at all. Now, let me ask the

model to do something a little bit more complex than just creating a static garden, which is to make me a Minecraft game. And on purpose, I'm not giving the model a lot of instructions. I want to see how good the model is at taking basic instructions like this and taking that and expanding on it based on its knowledge of what Minecraft is. So, let's see this. All right, looks like our game is done being made. So, let's just quickly look at what it thought about. So, okay, it recognizes that this is a fun challenge. It's going to use HTML, CSS, and JavaScript to make this and obviously 3. js for 3D rendering to make it more immersive. And it's going to add firsterson controls, voxalbased world generation, different block types. Okay, let's try this out. So, I'm going to click on click to play. Looks like I can move around. Uh, we can see that here. I can fly with F. So, let me go up. I can break stuff here. So it's not bad like adding some models even struggle to add the moving feature. So for me to able to do that I can do that and then I can change like what I'm trying to build. So like if I want to build this I can put that here as you can see. And then let me change to brick. So I think that is number eight. Yeah that is eight. So I can put some brick features here. So this is not bad at all like what I was able to create. As I said, some models nowadays struggle to even create like these movement features. So, it being able to do that shows how good this model is. And this is like obviously a much

cheaper model. Now, let me compare this model against Gemini 3 Flash. And the task I'm going to ask both of these models to do is make an interactive 3D scene with a big hollow sphere and a smaller ball trapped inside it. I can move or tilt the big sphere and the small ball must move only because of real physics. So you should understand gravity, friction, bounce, and collide with the inner wall. Give controls, basic UI sliders for friction, bounce, and mass. And make sure that the small ball never clips through the wall. So let's see this. All right, looks like both of these models are done creating. So let's look at Miniax M2. 1 first. So we have gravity here. We can increase or decrease the gravity. And based on that, the movement of the ball is going to change. Let's put the gravity all the way down. Let's just see what it looks like. Uh, it looks like the ball is moving a little against the wall, but like we can't really see much. Let's put the friction down as well. Does it change thing? Not really. The bounce. Let's put Let's increase the bounce. Let's increase or decrease the ball mass. This doesn't work. Rotation speed. Let's increase it. Makes it faster. But, you know, it's not really doing what I asked it to do. Like, it doesn't really look that realistic. Like, the physics are not really being applied. Let's just put on auto rotate. But the UI is not bad. Like we'll give it to the model like okay the UI interface is pretty good. But overall the model is not really like following the physics to a te. Let's compare this against a much stronger model which is the Gemini 3 flash. And let's see what it has created. So Gemini 3 flash. We can increase the bounciness here. friction and gravity. Let's just put gravity all the way at the bottom. Let's try this out. Drag the drag to rotate the auto superior. So, let me just do that. Uh, looks like it's moving, but I can't really tell. Let me increase the gravity. Okay, looks like Okay, this one definitely looks like it has better physics. Like, you can see that the ball is like following what it would look like. Let's increase the bounciness a little. Okay, now we can see that obviously this model is definitely going to be better. Like it's Gemini's probably strongest model and the flash model is pretty good. So compared to Miniax, obviously Miniax holds no chance, but the UI is pretty good. Like compared to both of these model, the UI is there. Like we don't obviously the same level of physics being applied here. Like over here, you have to use WD. So it works like the outer sphere looks like it's moving, but the inner ball is not. But I'm still gonna cut some slack to the model just because it's open source and it's cheaper and obviously it's not gonna beat Gemini. But the UI is good, so not bad at all. If you enjoy this

video, this is what we do here. Fast, clear updates on the biggest moves in AI. If you want to stay ahead of everything happening in this space, make sure you're subscribed. And if you want the hands-on side, demos, tools, workflows, and everything developers can actually build, will check out the world of AI. We also run a simple no noise newsletter that gives you the most important AI tools and updates in just a couple of minutes. Subscribe here. Follow World of AI. Join the newsletter.