China's Step 5 DESTROYS Gemini 3.8 Flash?

Julian Goldie (0:00)

We have a brand new model for free, which is Step 5 Preview. This is a brand new Chinese AI. And you can also use this for free for one week across open code, client, news research, Kilo code. We've already plugged it into our agentic operating system. So you can see that we can use Hermes Agent with Step 5 for free. Over here is a free model, which is pretty cool.

And then we can build with it. And then we also have this inside our free AI Coder. So we've been testing out as a free model inside here. And I'll talk you through what it is and how it works and why it's different, etc. Now, one thing to note here is it is not the most frontier model. So this is not going to replace Claude Opus or anything like that. Right. It's not really designed for that. But what it is pretty good for is a free model for sub agents. Also super fast. And you can use it as a free model for building and automating whatever you want in the background. Right. So you can see, for example, here we have the free AI Coder that we've built in. And then we can plug any new model that comes out. So previously we had LOX Alpha built into it, and we can use NIME router, OmniRoot, etc. with this. And then we can go inside the chat here and ask it to build something. Or we can go back to Hermes agent, and over here we can actually use it as a model. Now, this is pretty cool because we need really quick answers. Number one, Hermes Cloud is really good. But if we can use free models like Stepfun, that's really good as well. Bear in mind as well, Hermes Cloud is basically on a VBS. So when you're using it, that means that it's set up on a VBS. It's not affecting your local setup, but you can plug it into an agentic operating system like this. And you can use it side by side with something like Hermes Apollo, as you can see right here, that we've got ready to go, and also Hermes Oracle, and Hermes Music, et cetera. You could use it across all of these different workflows, which is pretty nice. So let's talk about what it is, how it works, why it's different, et cetera. So number one, this is 600 billion total parameters with 27 billion actives. It's a mixture of experts model, but also is 1 million token context window, and it has vision as well, so you can see things when you're using it. Now, it's performed on the benchmarks at a similar level to GPT-6 Luna on Hermes index. So if you're not familiar with Hermes index, this is a new index where they test all the models recently, and it scored 33.89. So this is not like GPT-6 Astra level, just to be clear here, but it is pretty good as a subagent model that works nice and fast basically. And you can see how it performs on the leaderboards over here. So step 5 preview.

It's around the same sort of level as HY4 and GPT-6 Luna, and it was built by StepFunAI.

Also, something interesting to know here is they seem to have skipped step 4 and gone straight from step 3.7 to step 5, and it can work with text and images, one million token context window, mainly text output if you're using this. You have performance on the benchmarks here as well. Now, if you're wondering, okay, where does this perform an artificial analysis index, which is pretty useful for comparing the models? You can see the is right in the middle, right? So it's a similar model to Luna or Gemini 3.8 Flash, which means it's fast, easy to use, pretty cheap even if you don't have it on the free model, and also it's outperforming a lot of models. You can see, for example, Mistral Medium 3.5 over here, Muse Glimmer, Cheapity 5.6 Luna.

These are older models, but at the same time, not that old, right? Like they just dropped last couple of months, really. So it's beating Gemini 3.8 Flash on both Intelligence and Cost, which is pretty crazy when you look at a Chinese AI lab, like one that most people have never heard of, and it's pretty much right up there with Gemini 3.8 Flash and outperforming it, which is wild. That's like Google's best, fastest model right now until Argon comes out. Now also one of the big differences versus Google, and this is a great article by the way by Office Char, you can see here that it costs 0.72 dollars per Intelligence index task. If you're using it on the normal API, bear in mind it's free for one week to test against 1.24 for Google's model. So it's about 42% less expensive to use something like this versus Gemini 3.8 Flash. But at the same time, it's outperforming it. So if you want to indicate how does it perform on Intelligence, you can see if we have a look at artificial intelligence analysis here, you've got step five preview over here, and then it's outperforming KimiK 3 Max and Gemini 3.8, and also DBCV 4.1. It's just behind Grok 4.6. Obviously, Grok 4.7 is out now, but you get the point, like it's performing pretty well on the tasks versus other models.