Ornith-1.0 is INSANE (FREE + Local + Open Source)!

**Julian Goldie** (0:00)
So today we're gonna be testing out Ornif 1.0, which is a new self-learning local model you can use. It scores pretty well on the benchmarks. You can see some examples right here. So this is a new free local model you can use. You can see how it compares versus Opus 4.7 here, and it's crushing on all the key benchmarks, as you can see. Now we've already plugged it into our agent operating system, which you can see right here, and we can switch the model anytime that we want. So we can also run this with Hermes too. So if we go to Hermes, we set up a separate profile for Ornif, and it can work agentically with Hermes as well, which is pretty cool. And you can get it for free on Ollama. So just make sure you have Ollama installed, and then you download Ollama, make sure you've got it open. And from here, you can plug this into whatever you want. And you can see how to get the model. Now, if you're wondering, how does this work? What is this doing, et cetera?
So basically, this is a powerful system where you can run a local model. It comes from a company called Deep Reinforce. They dropped Ornif, and it's an open coding model. This goes pretty well on the benchmarks.
If you want to see what we've built with it, here's some examples of different creations that we've made. You see the full gallery here. And these are all different things we've built with Ornif. So it actually can build useful stuff, interesting things, even apps and games and tools and that sort of thing, which is pretty cool. And some nice visual stuff as well. Right. So it actually works. You can plug it into Hermes as well. You've already seen inside the chat that we've got Hermes or Ornif plugged in, and it seems to work pretty well. Now, the interesting thing about this is, well, it's like you can have a self-learning agent, which is Hermes agent, and then you've got a self-learning model, which is Ornif, and then you've combined them together. Right, so you can kind of get the power of both just by using this system. If you're wondering how does it score on benchmarks, so it is outperforming Gemma 4 31B on the benchmarks here, and if we look how it compares versus flagship models, it's not doing too badly on there. I would test this stuff out yourself. I don't really pay much attention to benchmarks, but if you want to see the benchmark comparisons, you can see them right there. Now, what's interesting about this is self-improving. This is the strategy, this is the official blog from Deep Reinforce. So it's a self-improving training framework that jointly learns to solve tasks and to construct the scaffolds that guide those solutions. So it's building its own setup to get things done. So it'll roll out multiple solutions to solve the same problem, as you can see. So you're on it here, you give it the task, it starts working on building the system to build it out, and then it comes with three solutions to get that completed. And then from there, it self learns and improves so next time you use it. So rather than relying on a fixed human design harness, it treats the scaffold itself as a learnable object that co-evolves with the policy.
Basically what that means is it's self-improving and it figures itself out itself. It's a lot more independent. Now this runs with RL, which is reinforcement learning. And so the way this works is it gets the task, looks at what was previously used for similar tasks, looks at the task description, then generates a solution rollout.
And then that's repeated over training to give it a positive feedback loop so that it keeps improving over time. So if we look at the benchmarks here, for example, and we compare it versus QWEN 3.5, 3.7 max even, which is pretty impressive. I can't run QWEN 3.7 max locally, but if we look at the benchmarks here, we don't need to write on terminal bench 2.1 or NIF 1 is outperforming QWEN 3.7 max. Same for terminal bench 2.1 on Clawed Code, SW Bench Verified, like it's crushing on the benchmarks here. So if you're wondering, okay, how does the whole self reinforcement learning process work? So it writes its own instructions. That's the trick here. And it's a reason a 9B model that you can run on a laptop scores so well and better than models many times its size.
So most AI coders work inside a fixed recipe, a human wrote. The recipe, the harness tells the model how to solve a problem. Look at first, try this, then try that, etc. And every model gets the same recipe. But if that recipe is wrong for your problem, the model is stuck with it. Whereas with this situation, Ornif would write its own recipe, which means for each task, it does two things. First, it writes a plan, so how it's going to solve the problem. And then it follows our plan to write the actual code. But when the result gets graded, the score doesn't just teach a code, it teaches a plan too. So both halves get improved.