Engineers, it's official. Opus has real competition. GLM 5.2 is the leading open weights model, and right behind it is another really fantastic model, Minimax M3. Normally, I ignore the open weight models. In the past, they haven't been good. They haven't been competitive. That is no longer the case. It is now official. Opus has real competition. It has competition on multiple dimensions. Of course, as soon as Fable is launched, that competition is going to disappear, but there are other important key pieces of information I want to share with you here that can help your agentic engineering. So, let's frame our focus here on GLM 5.2 versus Minimax M3, because this is going to reveal four big ideas and 10 insights. Here they are. If these ideas interest you, stick around. Let's jump right into the conversation. The headline is simple. Opus has competition, but it goes deeper than that. As you know, the industry is undergoing a massive shift right now, and one of the big things I'm worried
[00:01:00] about, one of the big things many engineers are concerned about right now, is can an AI lab, the government, or anyone kill your agents at any point in time? We're thinking about resiliency. We're thinking about control and true ownership over our AI. And this is where open weight models really stand out. There are many providers for GLM 5.2, for Minimax M3, and we're going to discuss when and how can we really own a powerful model like GLM 5.2 that is near the front tier. This is not a benchmarking video. I want to give you the insights. I want to give you the oil of all the benchmarks out there, of all the information up there, so you can really understand the open weight landscape in a very efficient engineering first way. So, this is not about live coding. This is about engineering. This is about building out your engineering agents and your product agents. And then, at the end, I'm going to share my model stack. So, I'm going to share how I'm organizing all of these models on three tiers, [music] three categories, which we'll break down right now, State-of-the-art workhorse and lightweight models. So, of course, we're
[00:02:00] going to have the cracked Fable in there. You already know where that's going, but let's refocus on GLM 5.2 versus [music] Minimax M3. What I like to do with all of my models is I place them on one of three tiers. We have the state-of-the-art tier, we have the workhorse tier, and we have the lightweight tier. This is a quick mental model for understanding where the language models capability sit. We're going to use Opus as our max control group. GLM 5.2 and Minimax are going to be the experimental groups we're going to be focusing all of our attention on. And then we're going to also have a min control model. Comparing a model in isolation is useless. So, what I like to do is I like to set up a state-of-the-art comparison. So, as you can imagine, Opus 4.8, this is the ceiling. This is the very top of the capability of models. There are others in this category, but just for this experiment, just to compare GLM 5.2 and Minimax M3, we're going to have a single front tier model. And we're also going to compare it against Quinn 3.6. So, this is running
[00:03:03] right on your local hardware. This is a lightweight model that you can own right now. So, when we have a max control model and a minimum control model, this gives us a clear framing to inspect and understand GLM 5.2 and Minimax at a deeper level. GLM 5.2 is going in the A tier and Minimax M3 is going in the B tier on raw performance. This is a performance benchmark, by the way, just to make that super, super clear. This is how I like to organize my models. I think about tiers of models, three tiers, and a single stack. So, it's not about picking one model, it's about having multiple options for your engineering agents and your product agents. So, here's the headline. I'm just going to give it all away right now. GLM 5.2 is the better model, but Minimax M3 is the better deal. So, the only real question that should be guiding your decision on which model to be using here, again for engineering agents and product agents is this. Do you need maximum capability or can you optimize for price? And that basically will flip your tier list. Opus goes
[00:04:00] right to the bottom, GLM goes right to the B tier, Minimax goes into A tier, and Quinn, of course, jumps into S tier. You can run this, air quotes, for free on your local hardware. And with a cloud provider, Quinn becomes a lot more powerful and fast while paying a very minimal price. So, that's the headline. GLM wins on performance, Minimax wins on price. And this is where things get pretty interesting. So, if we look at the raw stats of this, this is what really matters here. It's performance, speed, and cost. So, it's very clear Opus is in a decent lead above these open-weight models. A lot of hype-centered content creators and creators on X, they're going to tell you that GLM is a replacement for Opus 4.8. And as you likely know, that is not true. They are competitive. They're competitive in two dimensions, performance and cost, but they are not going to beat Opus. Let me just de-hype that. GLM cannot replace Opus. Let's be super clear about that. But both GLM 5.2 and Minimax have some pretty great things to offer, great speed, and of course, great price. Now, this is where things really break down and where the
[00:05:00] numbers don't look so good for Opus. Starting with tokens per second, as you can guess, these lightweight sub-100 billion parameter models, they're very, very fast and they're nowhere near as fast as they can be when we start adding things like speculative decoding, multi-token prediction. There are many ways to speed these models up even further. If you're on a Mac device, MLX is going to be a great first place to go. GLM's surprisingly fast, but the catch with this model is that GLM 5.2 thinks a lot. Many, many, many of its tokens are just reasoning. And this shows up directly in the benchmarks. So, on artificial analysis, if we scroll down to output tokens here, GLM 5.2, if we look at the total output tokens per task, pretty much all the tokens were spent reasoning. And it's even more than Opus in max mode. It's much more than Minimax M3. So, very interesting here that a model can be quite fast, but if that speed is put toward thinking, it doesn't really, really matter. Because what we're looking for is response time.
[00:06:02] It's that total wall time that matters when we're thinking about getting results, when we're thinking about calling tools, and when we're thinking about the experiences that users will have when they're using your product agent. Okay, so inside one speed, all these models are viable. All these models are usable. This isn't really the differentiation point of these models. So, insight number two is the raw comparison. So, what is the differentiation? It is, of course, performance against price. So, this is what really matters in the, you know, the decision tree is quite simple. You should pick GLM 5.2 when capability is the most important thing at a lower price than Opus. And then you're going to want to pick Minimax when cost and volume are the constraints. And, of course, when Minimax can do the job. And this is like always the decision tree here. If the model can't do the job, it doesn't matter the price, it doesn't matter the speed, you can't use it. And that's the interesting part about GLM 5.2 is that we're getting into the space where GLM 5.2 is performing on par and
[00:07:00] slightly under a lot of the tasks Opus 4.8 is executing on. But when it comes to GLM 5.2 versus Minimax, this is the big takeaway. Use GLM if you need performance, use Minimax, though, if it can do the job. This is where we get to the cost curve. Because things get really, really interesting when you just look at the cost curve. The number here is that every time you drop a tier of capability, and the tier capability is becoming smaller and smaller every single month. But every time you drop a tier capability, you basically drop your price 5x. And it's the same from GLM to Minimax, and it's the same from Minimax down to Qwen 3.6. Although at this level, you lose too much capability. The Qwen models, although very powerful, although on device, although private, which are all important things we're paying close attention to, the intelligence must be at a certain level. I got to say though, you know, the more I experiment with Qwen 3.6 35 billion reactive, the more I like it. >> [laughter] >> And the more excited I am for the next Qwen local model release. The cost curve looks terrible. Like it's a cliff. It's
[00:08:01] a 5x cliff. And again, you just drop down one tier. And with each tier capability, you're not losing that much. Okay, so that's insight number three. The cost curve looks really, really rough for Opus. Opus 4.8 doesn't have much longer on the market before open weights models catch up pretty indefinitely. With that being said, insight four is us being really balanced here. Engineering is all about trade-offs. We'll talk about that more in a moment. The workhorse models called tools like Opus, but they don't ship like Opus. You can look across every single benchmark in this benchmark here on artificial analysis. I've dialed into these four models to simplify the comparison. But if we just search GLM 5.2 max, across the board this model will always lose to Opus. And sometimes very dramatically. And you can see here, every once in a while it's losing to Minimax as well. But the thing to pay attention to here isn't just all the benchmarks that GLM 5.2, it's the emergent behavior that emerges as you go up on these benchmark
[00:09:02] scores. Like all of these benchmarks are proxies for some capability. Terminal bench, super popular one, agenda coding and terminal use. This is just one proxy for agenda coding. The benchmarks do a decent job alone. When you really put together the index, it means a lot more. And so when you just look at the raw numbers here on how much better this five points looks like a small amount, it's not. It's a massive amount. The five points of difference on the index means a lot. Just because you can call tool doesn't mean you're a great model. And things really start to fall off for long horizon tasks. Whatever Anthropic is doing to train their Claude series models, specifically Opus and Fable. This is a huge point of differentiation. Workhorse models called tools like Opus, but they don't ship like Opus. The results are not the same. So, there's still a gap. Let's be super honest about that. GLM is not a replacement for Opus 4.8. It's a decent competitor. I always think about this trade-off
[00:10:00] triangle. As I'm working through benchmarks, as I'm creating my own custom benchmarks, and as I'm building engineering agents and product agents, I'm always comparing these models with this framework. And this is how I like to visualize this. There are three axes. You have performance, you have speed, and you have cost. And in reality, you get to pick two, never all three. Very clearly, Opus gives you performance. Minimax is going to give you cost and performance, more on the cost side. Quinn, of course, this is all speed and cost, probably more on the cost side. And then, GLM is going to give you the speed and performance. And honestly, this one's pretty good on cost, as well. GLM's probably going to be that closest one that does all three for you. Every single model has a trade-off. And if we're going to be objective, if we're going to do actual engineering and not vibing, we need to be honest about where all these models are. I use this trade-off triangle all the time to help me place these models. I use this and the tier list. Engineering is all about trade-offs. Language models and agents are no different. Every model can be placed on this chart. We can probably push GLM's performance up a little bit
[00:11:00] more here, as well as Minimax. But this is really what it's about. Understanding your model and then creating a model stack, which we'll talk about more later. I keep saying engineering agents and product agents. Let me just clarify this. Engineering agents are the first version of agents. Thanks to Claude Code, thanks to the first Claude series models, agents for engineers have been unlocked. But this is just a fraction, literally less than 1% of all the agents that will be created in the future. What are the other agents? It's product agents. Here, I'm just grouping every other domain under product agents. And these are the agents you'll deploy into your products. This is where tokenomics really matters and where you have to arbitrage your tokens because that's the business. For engineering, we can be a little bit more loose and dynamic with our token spend because of course we're still in the experimentation phase, but it's also harder to measure the ROI right now out of engineering agents. All that being said, it's still important to know how and when to use the right language model for your agents. The simple guideline is if you can afford
[00:12:01] the best compute, use it because you'll be getting the best results. At the same time, as these models catch up, as we discussed in our Fable band video, just like Opus can do a lot of what Fable can do, we're getting to that point where some of these open weight models can do 80, 90% of what Opus can do. And I'm specifically looking at GLM 5.2. If you've been using this model, you understand that it is getting close. It is competitive with Opus 4.8 at a fifth of the price. It's becoming harder and harder to ignore that. So, again, just thinking about that three-tier framework, right? We have state-of-the-art for our hardest work, and then we have daily drivers and high-volume agents. These are your workhorse agents that can do a lot of work, not at the best performance, but it's going to be a lot cheaper, and by a lot, we're talking 5x cheaper, and then 20x cheaper if you drop down to the Minimax tier level. And then of course, we have what we're all looking for, local private models. This is coming. We need more time. As mentioned, we're going to talk about some of the hardware requirements needed to ship and really use a GLM level model on local hardware
[00:13:00] in just a moment here. So, this is how I think about engineering agents. Basically, if you can, spend all the compute. If you can't, delegate to smaller, cheaper models. And I think it's important to understand the local model capabilities of models like Qwen 3.6, Gemma 4, so you really know what is possible on device right now. I'm going to be creating a lot more content around these local models because they're catching up and there are a lot of use cases that they can cover now while being local and private. So, make sure you like, make sure you subscribe, all that good stuff. Let's keep rolling. Product agents is a bit different because tokens matter. The tokenomics matters. Routing to the cheapest model that clears the bar, that satisfies your users is what makes or breaks your business when we're talking about your product agents. So, for example, with our two models we're looking at right here, GLM versus Minimax, when you have users at scale and we're talking tens of thousands, hundreds of thousands of users, you have to think about the tokenomics. You cannot just throw Opus at it. It's not scalable. As you can
[00:14:01] very easily imagine, at the same time, something like Qwen 1.5 6 super low performance, all that to say very, very simple task. And I'm not talking about role-playing, I'm not talking about these stupid really low-level entry use cases for these models. I'm talking about like real work, right? We're talking about accounting, we're talking about finance, we're talking about healthcare, we're talking about biomedicine, we're talking about getting value out of these models at consumer level scale. These models make businesses run now. GLM 5.2, Minimax M3. If you study these models, if you understand the tokenomics, if you understand the performance, you know, GLM is your performance per action and then Minimax M3 is going to be your cost per action winner here. This matters a lot for your product agents and I'm building out products right now that I'm consciously making this decision on. Do I need performance here? Do I need cost here? When I benchmark both these, what does the result look like? When can I use Minimax? When should I use GLM 5.2? Another important insight I want to share with you, three of the four of
[00:15:00] these models can't be switched off. We're all up to date on what's going on with Fable, the government, Anthropic. Insert any AI lab substitutability isn't a footnote, it's the whole strategy in 2026 and beyond. If you're going to scale your engineering agents and your product agents into production, we can't have models that can just switch off at any second. Now, of course, there's a lot of fear around this. The economics around shutting down an AI model is so bad for Anthropic, for OpenAI, for Gemini, that they would never allow it to happen for too long. But, um things that can happen do happen. Right? As we saw with the Fable model. And it's not like mythos class of the models are going to stop. It's being said that Anthropic has finished training the model after Fable. So, it's not like the model progression stops. But, the key here is we're not in control of this. This is a closed source model. Fable was the last straw. We cannot fully depend on the most important technology to be available to us. We must make sure it's available to us. And so, these open models are going to be really, really important for that. Right now, only lightweight models are
[00:16:01] available to really own end-to-end. I definitely recommend, if you don't, grab whatever chips, any device you can, so that you can get your hands on these sub-50 billion parameter models with the mixture of experts on it, so that you can really have local AI, no matter what. But, what's coming next, and what I think a lot of us are really excited for, is when we can run GLM 5.2 and MinimaX M3 on some local hardware, on our home lab. Now, this is possible now. Some prosumers, some engineers, you know, spending tens of thousands of dollars can do this now. So, let's analyze like what does it really take to own GLM 5.2, a powerful workhorse model. GLM is available right now. You and I can own this, but it's just not all too realistic yet when you look at the actual cost breakdown. So, an M5 Max MacBook Pro device like mine, I cannot run this. Okay, so I'm out of the running already. I need custom hardware, custom chips, Nvidia hardware to be able to run this thing. At about a budget of
[00:17:01] 2 to 4,000, you could probably get a serious 1 2-bit quant of GLM 5.2 running on your budget home lab at a very slow 6 to 11 tokens per second. To me, this is unrunnable. So, this is not worth it at all. Now, at the next tier up, things become a bit more viable, but it's going to cost you. This is usable, 10 to third tokens, but you're going to need something like Mac Studio M3 Ultras or DGX Sparks stacked up. Bandwidth is going to be the constraint here. Again, this is viable. This is something you can do right now. It's getting harder and harder to find access to RAM and it's becoming more and more expensive, but this is available. Now, where this really becomes viable is if you're willing to drop, you know, 50 to 90k, really 50 to 100,000 dollars on, you know, something like six RTX Pro Black Wells. That's going to give you a 500 GB of RAM. Then you'll be able to run the like 4-bit quant GLM 5.2. As you can see, very, very expensive. I have a rough timeline here. This is just my rough estimate. I'm looking for that mid-2027
[00:18:00] time period to really get a GLM class model running locally on my hardware, on some box. I'm really hoping for that next gen M5 Ultra from Apple, but if they don't put it out, I'll build it myself with some Nvidia and AMD hardware. Whatever I have to do to get that, it's now a priority for me to make sure I de-leverage off of closed-source AI models. As we kind of move toward that place, you know, not all of us are going to drop tens of thousands of dollars on compute when we can take this compute and throw it against an API. So, you know, I just want to list out the options for you. This is how I'm thinking about this right now. There's kind of four options, right? You can set up a home lab. You can rent GPU by the hour. So, if you're running a business at scale, this is an option. Someone else can host the open weights. There are tons of providers. I think this is going to be the current action for most of us because we can always have one of 20 providers up and running for us. Very likely you'll only need two or three, but you for sure want to de-leverage off of a single provider for the model that's running your business. Specifically talking about product agents here, but also engineering agents
[00:19:00] as well. And then there's of course just paying per second, scale to zero, best for burst to uses. But this sits very closely to hosted open weight APIs. So, these are kind of the options that I'm looking at. I'm really, really looking forward to getting the home lab, but it's not all too realistic. And it's going to be a lot simpler, especially if you're running a business, to rent GPU by the hour or to do what we're all doing right now, which is just to hit serverless. So, the story for owning your own GLM 5.2 is basically you have to wait a year. Right now, we're stuck using cloud providers. But that's okay, because we have more cloud providers than ever. GLM is hosted on some 10, 20 different providers. I'm really, really looking forward to this. There is a nightmare scenario here where GPU and and RAM prices continue to just go up throughout 2027. This is a real possibility, which will then push out this timeline even further for owning these workhorse models locally on your device. So, I think Minimax is smaller, right? Minimax is 400 billion parameters and runs 23. So, Minimax is a lot closer to being run on some local hardware. If
[00:20:00] you really want that Opus level performance, we're going to have to wait a bit longer. But this is just something to keep your eyes on. Don't pick a model, pick a model stack. So, this is one of the, you know, big ideas that I've been talking about on the channel more and more and more. You don't want to be dependent on one model. You don't want to be dependent on one model provider. Everything going on with Anthropic, with Fable, with the government is kind of showing us that right now. It's time to pick a model stack. Let's elevate ourselves. Let's de-leverage. Let's become more resilient as agentic engineers. We need model stacks. So, these are all the models I'm actively using, researching, experimenting, and using in both engineering and product agents right now. I wanted to kind of share my tier list with you to kind of give you an idea of where these models belong in each category. Let's start with the top and the bottom. So, Fable 5 is S+ tier. No other model is going to be an S+ tier. This is the best of the best. As soon as this comes back online, I have several gnarly experiments that I want to run, and I'll be doing heavy, heavy multi-agent orchestration. Check out the previous couple videos where we talked about this model in detail. So, that's
[00:21:00] our very state-of-the-art. And then at the bottom, we have our Gemma, and we have the Quint 3.6 35 billion parameter model. So, I'm actively using these, testing, building on these. These are the lightweight models. At the lightweight level, you own this fully. That's one of the properties of lightweight, it can run on your device. So, we're talking sub 100 billion parameter model. Then we have our workhorses. So, these are the models that could run up to 90% of what you're probably using state-of-the-art for. These models are the GLM 5.2 and the Minimax M3. And the only difference here really between the A and B tier is a performance hit of maybe 5 to 10 points on artificial analysis. GLM 5.2 down to Minimax, it's going to be like 5-6 points in that 40 to 50 range. And that's going to be pretty consistent across every single benchmark. You're going to see a 5 to 10% drop in performance by dropping down the A tier to the B tier, but you're still going to have a great workhorse. And with great prompt, context, and harness engineering, you can easily make a B tier an A tier. But, you know, by
[00:22:00] default, out the box, your B tier is going to be a little more optimized for price, while your A tier is going to be more optimized for performance. I'm just going to go ahead and place the rest of these, right? These are all the models I'm using right now in my model stacks. Opus, obviously S tier. GPT 5.5 coming right before it. We have a ton of workhorse models here. So, Gemini Flash coming in here. We have Gemini 3.1 Pro. I'm going to throw that behind GLM 5.2. We of course have our Deep Seek Pro. And then we have Deep Seek Flash. That's going to be your B tier model. Uh Kimika 2.6, that's going to be a B tier. And then Quinmax, this is also going to be an A tier model. This is my model stack. And the whole idea here is you want to think about your models in tiers, in a stack. You want to have your state-of-the-art for your hardest work, for your engineering work. You want to have your workhorses that you put in your products. And then you have your lightweight that you can run locally and use for private personal things. And eventually, the idea is that all these models over time will move down the stack while new ones come in to the state-of-the-art and in to the S tier. And this gives you and I more and more control, but we have to choose the right
[00:23:02] model for the job trading off performance, speed, and cost when it matters the most. And a great side effect of this is that you deleverage off of the closed sourced models that at any point in time can just rug pull or be rug pulled as we saw with the Fable model. We have a ton of open weights options now. I think GLM is actually outperforming 3.5 Flash. We can go ahead and look at that. Exactly, yeah, it is. So, it's just a point above Gemini 3.5 Flash, which is pretty insane. Minimax and GLM, two really, really great models. It's a simple question of performance versus cost. It's incredible to see an open weights model this high up in the benchmarks. I don't think an open weights model has ever hit this high. This is a top five model, which is a, you know, pretty insane thing to say. If you're looking at pure intelligence, GLM 5.2 is top five. And also, pretty insane, Minimax M3 is also top five, although it's on the edge here. There are other models that can do what Minimax can do. For instance, DeepSeek V4 Pro is a great example. As mentioned,
[00:24:00] GLM 5.2 is also very fast, but most of those tokens are going to be reasoning tokens. So, it's not going to really feel fast unless you're staring at the thinking trace. And as a side effect of that, the GLM 5.2 price is actually going to be a little higher than you want it to be. Although, it's going to be a decent amount cheaper than your top end models, of course, like the GPTs and the Fable fives. I'm personally a really huge fan of Minimax 3. I've been enjoying using this model and really dialing in the system prompt of agents running this model so that it runs in exact straight lines at really, really, really cheap prices. This is a capability that I think is super underrated right now. If you really specialize your model, uh if you build agent experts as we talk about and discuss in Agent Horizon, your model can really, really bump up in capability, but that's if you're being specific and if you know the exact repeat task that you want your Minimax, your Flashes, and your Kimmys to execute. And really all of your workhorse models. If you can really guide and steer your workhorse
[00:25:01] models, you can drop down your price even further. That's where great agentic engineering comes into play. So, this is my model stack. This is how I'm thinking about great open weight models like GLM 5.2 and Minimax M3. If you got value out of this video, if the model stack framing helps you engineer, drop a like, subscribe, join the journey. We're on a mission to build software that works while we sleep. Comment down below. Let me know what your model stack is if you have one. And if you don't, comment down below and let me know what you think your model stack is going to be. Building your model stack is going to be more important every single day as we use these powerful fable mythos class level models to orchestrate other work and as we build out individual agents running inside of products for our users. We're in the age of agents. So, knowing the best tools, models, contexts, and prompts to equip your agents with is the name of the game. And when you have optionality and resiliency built into your model stack, when one of these models go down, you will be the
[00:26:01] engineer unaffected continuing to win, continuing to ship as everyone moves more and more into the vibes of all these models. Let's stay focused on great engineering patterns and principles. You know where to find me every single Monday. Stay focused and keep building.