Ling 3.0: New FREE Chinese AI!

**Julian Goldie** (0:00)
Today, we're gonna run through Ling 3 Flash, which is a new meter of experts, model 124 billion parameters, 5 billion active per token. This is a new free API you can get access to on OpenRouter. And the good thing about this is that it's actually performing pretty well on the benchmarks. I'll show you what we created in a second, if you wanna see how it performs. So in terms of benchmarks here, you can see it's actually outperforming DeepSeq V4 Flash again. We've tested it out. I'll show you some examples in a second. And it's also crushing ChapGPT on some of the benchmarks like SWE Multilingual, Terminal Bench, and Wide Search. Again, you always wanna test stuff out yourself. Don't just believe all the benchmarks.
You can get access on OpenRouter for free. You can also get free access via Hermes, which I'll come on to in a second too. And this is the new free model. Now, if we have a look inside the apps in terms of what is being used the most with, Sequila code, Hermes, Agent, Chord code, OpenCore, and Pi, are all being actively used with Ling 3.0. So it is a very popular model. And when we actually test it outside the chat, it's not bad at all. So if we go to the chat side of the router here, you can just select the model you want to use. And here's the HTML. The one thing that I noticed is that it's a lot faster. It's also 250 context window, which you might think is a bit low, or you might even think like the 124 billion parameter is pretty low.
So Ling have actually come out and said with one-eighth of the total and one-twelfth of the active parameters, it matches or beats their one trillion flagship model on most benchmarks. They've also demoed some examples with Blender MCP, which we'll show in a second. So let's see how it performs on the benchmarks here. So this is a website where she created with a one-shot prompt. Super basic, just said, create a beautiful website for an SEO agency. And you see here, it's actually pretty decent. I've seen a lot worse from free models. Wasn't expecting it to be as good as that. And bear in mind, it's just one shot. So you can change it, you can tweak it, you can switch however you want. The links at the top actually work. I mean, it is using those generic sort of place markers for testimonials and that sort of thing. One thing that would be pretty good is like it can actually code. It can do a decent job on front end. What you can do is give it a skill like hallmark on. This is like an anti-AI slot design skill, and then it could actually create like pretty nice designs for you, I think. We have a look at some of these examples it's created. You can see here that it can actually connect to MCPs. So you could train it to use an MCP like Blender, and then it can create like 3D models for you. It's also supposed to be pretty good at research, office work, and that sort of thing. We can also compare it side by side versus chat chibity. So let's try that right now.
We can compare the output side by side in a second. If you're wondering how do you use it with something like Hermes Agent, so you can go over to the terminal here, then we can select Hermes model. From here, you just want to select News Portal, log in like so, just switch to the free plan. Now we're connected on the free plan. And then if we go back to the terminal here, we can choose one of the free models. It's actually a bunch of free models, but we're going to go with Ling 3 Flash. Then we'll run that. And you can see it's actually working now with Ling 3 Flash for free.
So if we just do a random example, like forward slash learn, and then we'll give it an example guide. Let's try this one. So let's run a test command right here. Just had to update Hermes. I'm on a separate laptop right now. And now you can see it's fetching the guide, and it's using learn to help us. So a good thing as well, like it's a faster API, it's pretty quick to respond. You can also use it inside Kilo code.
So if you scroll down here, you can select link 3 flash, and start giving Kilo code an app idea to build. So let's try this now. And if we come back to Hermes agent here, you can see that it's created the skill with the forward slash learn feature. That was pretty quick. Actually works pretty well. So the app is being built inside Kilo code as well. And whilst that's being done, let's go over to open root. And this is the output from GPT 5.6 Sol, from, let's test out the one from Link Flash and compare them side by side. So if we add a new habit, this is a Link 3 version. Assess this out, select a category. So the app actually works. It's not bad. I mean, I would still say, of course, GPT 5.6 Sol's version looks nice. So let's test if it actually works. So this one looks a bit nicer on the front end, but they're both functional, and there's not a huge difference between them. Both the menus work, the tracking works. It's a fully functional app, built in one shot. We've also got the output back from KiloCode for the app. So this is built for free with LingQ 2 And if we test this out, Spaghetti 5000 calories, Boom Shack, Alaka, or actually worked.