Introducing Ornith 1.0 - Agentic Coding LLMs

Okay, so just as news of GPT-5.6 has come out that the government is going to be preventing people to use this, open weights models keep chugging along. And this brings me to the topic of today's video, Ornith 1.0, which is a family of models from Deep Reinforce. And I guess the big tease here is that things have been moving along where we've had sort of one track B models and one track B harnesses. And what Ornith is proposing is why not have a model that writes its own harness, or at least has the ability to write its own harness. Therefore, if you've got a particular use case that needs one kind of harness, it can write that on the fly and then be able to use that to actually get the result that you're trying to get. And this is exactly what the Ornith family of models are trying to do. They actually call this self-scaffolding LLMs for agentic coding. Now, the area that they're focusing on is agentic coding in here. I think the lessons can probably apply to other domains as well, but it does seem the whole topic of agentic coding is where the money is at the moment and where a lot of people are trying to make the most innovations. Which makes sense with everybody using things like Claude Code, Codex, and all the other agentic coding tools that are out there. So, that brings us to what is actually in Ornith 1.0, their family of models. Well, this is four models that are not new pre-trains or anything. These are basically fine-tunes or sort of more mid-training and post-training on four separate models that come from the Qwen 3.5 family and the Gemma 4 family. So, we've got a 9B model, which is Qwen 3.5. We've got the 31B, which comes from the Gemma 4 family. And then we've got a 35B MoE and a 397B MoE, both from the Qwen 3.5 family. Now, the cool thing in here is that they've released all of the models, unlike some of the things we've looked at recently where they've held back the biggest and best model. Here, we've got all of the models available, so you can try it out whether you're just looking out to try the 9B small model, whether you want to try the 35B MOE, or whether you want to go really big for the 397B. But, the key thing that sets this family apart is that these models are learning to generate both sort of solutions roll out, so agentic trajectories, and task-specific harnesses that guide those roll outs. So, their whole goal is by optimizing both that sort of scaffold {slash} harness and the solution that the model is then going to basically generate better solutions out. And if we just quickly look at the benchmarks, I'm not going to harp on about these, but certainly looking at their really big model here, we can see that it's outperforming a lot of other models, including the Quen 3.7 max, the mini max model, and often is competitive with the Claude Opus models. If we do the same thing and look at Even the smaller ones are doing very well against models their own size. And that's for the 35B MOE and even down to the 9B. So, the 9B is certainly something you should consider if you're looking to run a local coding model and you can't run something as big as the 35B or or the Gemma 431, etc., their 9B model is doing really well here compared to those other models that are often even three times as big as them. But, what's far more interesting than the benchmarks for me is what actually makes this special. And this goes to the heart of sort of what they're getting at here is that Up until now, most of the time people thought about agent harnesses as being things that are designed by humans. So, what Oracles actually does here is they treat that as a learnable object. And the whole idea is that just as the model's going to learn to be able to get good results, it should also learn to create better scaffolds or better harnesses for creating those good results. You could even think about it as their model learning to do context engineering instead of you having to write that by hand. And that idea here of how do you actually train a model or sort of mid-train and post-train a model to be able to do this without it all falling apart is the key thing to me about this project. And this really comes down to this sort of two-stage IRL process that they've got here. So, they basically propose a task, and then they condition on that task and a scaffold previously used for it, and then the model will first propose a new sort of refined version of that scaffold or that's harness, and then conditioning on that new harness, it will then basically propose the rollout which is going to get us to the result that we want. Now, then what they do is they take that rollout and because they'll have multiple rollouts there, and they will then basically use that as a reward signal to update the model for both generating the scaffold and generating the rollout here. Ideally, what's going on here are those reward signals, and this is coming from using GRPO here, and I'll talk about how they're doing the verification in a minute, but that allows them to basically update the weights of the model to get better at building a scaffold and then better at using that scaffold to actually generate good results out. Now, currently, they're focused on doing all of this in the domain of agentic coding, so it kind of makes sense, and based on their benchmarks, it seems to be working for that. But, you could imagine this being used for other kinds of tasks. Now, obviously, the challenge that you've got is can that reward signal be something that you can measure with GRPO, is it something that's verifiable, etc. And the thing you should be thinking about as you're sort of looking at this is, well, if the model is determining basically how the harness works and the actual rollout, why doesn't it just cheat and build a harness that just takes a shortcut to get to the right answer? And obviously, this is something that they've thought about a lot, too. They actually address this in how they deal with reward hacking. So, remember reward hacking is where if you've got some kind of LLM as a judge or something like that, the model just learns to do tweaks that are perhaps not exactly what you want, but that are going to get the best reward out of that model. And in the early days of RLHF and RLAIF, etc., you would find that the sort of generator models would generate all these kind of weird ways to get the reward model to just give a high signal back. And that could just be repeating certain tokens, doing stuff things that as humans we would clearly see as being wrong, but because the reward models were perhaps not as good back then, they could be tricked quite easily. Now, that said, even though the reward models nowadays are better, there's no guarantee that that alone is going to give you the best reward signal to actually update your model. So, they talk about here of basically defending against this in three layers. So, the first one is the actual environment and sort of tools and I guess basically everything that's going on in the sandbox that the scaffolding is going to run in. So, all of those things are immutable. They can't be changed by the model. So, the model can't set up the environment to actually do a shortcut to sort of win kind of thing. Secondly, they've got a deterministic monitor here, which is kind of cool, right? That what they've got is something that's basically watching what does the scaffolding actually try to do? Does it try and do anything like modify verification scripts, use any sort of unsanctioned tools, try to use anything that's out of the sandbox that it's allowed to play in. And the moment that it does that, it basically gets penalized for doing that. And the third thing that they have is an LM as a judge that can veto on top of any of the other stuff. So, even if it passes the early stages and stuff like that, if the LM as a judge also says no, this has been achieved in some way that perhaps wasn't allowed, it can then veto the whole verifier that the GRPO was using, etc. So, overall, this is a really interesting idea here of can a model then actually sort of code what it needs as a harness on the fly. And we've seen the idea of models actually writing code that they can then use be around for quite a while. So, this is the PAL paper. This is from It was late 2022 was the first version of it. This published version is January 2023. And what was so cool about this was that it basically proposed that why doesn't the model just write Python to do things like math and things like that and then run that Python and bring that back into the model. So, that kind of idea is actually quite old. But getting it to write a whole scaffold shows that one, not only have we come so far in the models that the models have gotten so good at this, but this could be the way forward that as the models get more intelligent, we don't want to basically get involved as humans much at all. We just want to leave it to the model to write what it needs to get the result that we want out. So, let's have a look at the model, see what the actual outputs look like, and get a sense of how this is all coming together. All right, so if I come in here, I'm using my testing suite in here. You can see we've got the model. I'm using the 35B MLE here. If I give it something like this, right? So, just starting off with some sort of general things. Draw pelican. So, the whole pelican SVG test, it draws a pretty nice pelican in here, right? So, you can see that like this is straightaway getting a pretty good result for the pelican, which is kind of be expected. I'm guessing that I can't remember testing the Quen 3.5 model on this, but that does quite well. If we give it a rag question, you can see we're getting it where it's writing up a lot of sort of reasoning thinking tokens in here. It's basically gone through and looked at the different stuff in there, but it hasn't really written any code or anything for this. It's just worked out, "Okay, this is what should be the answer." And then it comes back and gives us answers in there. Again, still pretty impressive. And it does that for quite a number of the other sort of tasks as well. If we give it something that where we say, "Create a harness to get the weather with a 5-day forecast." And we run this. By the way, the computer here is sponsored by Dell. I'm actually running this on the Dell Pro Max with an RTX 6000 Pro here. So, I'm getting very good speeds that you can see for the model. I'm able to run both the 16-bit models and also quantized models here. This one is actually a quantized version that we're running through a llama. Uh you can see for some of the other ones, I'll run them through VLM directly in here. Anyway, if we come in here and look at the thinking process, we can see that it understands the goal of create a harness. And it understands that it's going to need to handle all the different things that the harness does. So, this is what is kind of unique with this model. It then basically starts to design a solution. We're getting very nice long chain of thought here, which is very clearly related to the task at hand, going through this. And then it starts to draft the code. In this case it's done 1,600 tokens in there. If we come in here, we can see we've got it basically deciding it needed to have the request library. It then basically creates the weather harness. The weather harness is going to be for 5-day forecast in here. We can see that it's decided how it's going to handle a bunch of things for this harness and to basically give us some graphical displays, etc. And in this case it's telling us, "Okay, we need an open weather API key here." And it shows us what some of the outputs should be like, etc. Okay, if I change it to basically say, "Right, we don't have any API keys." Let's see, do we get something now like an API which doesn't need an API key in here. So, it's still generating a lot of thinking tokens. My guess is that the thinking tokens going to be very similar. Although, we see the implicit need now is find a free no API required weather source. It's gone and generated a lot more thinking tokens as it's gone through this. And it's come up with the result of using the open meteo API, which is this API here. Sure enough, doesn't require a sign-up in here. So, it's then basically rewritten the script to match that API. So, we can see we've got these codes here, weather interpretation codes, and it will convert them to emojis and to a sort of short description going through this. Then it's going to take coordinates as GPS coordinates. It's actually got the API in there. It looks like it's handled all of that very well, which is really good cuz it understood the requirements that I don't have any API keys and had to basically do it without doing the API, etc. And you can see now if we ask it something like create a harness to get the latest AI news from Twitter. Let's see what it's going to come up with. It is really nice that when the tokens are this fast, you don't mind having a few thousand thinking tokens going through this. And we can see coming in this, okay, get the latest AI news from Twitter's API has changed significantly. It's doing quite well at understanding that. It's then got code for it dealing with the actual sort of main Twitter API. Okay, what if I change this to telling it I don't have any API, so find a solution that doesn't need one. Are we going to get this in the thing? Okay, no API keys allowed. Need a solution that bypasses this requirement. So, you can see that the chain of thought has kind of been trained for very specific things. First, understand the user request. So, this is a common thing that it is in the sort of hidden chain of thought for a lot of the proprietary models. It's got the constraints there. It's got identify key challenges, constraints, explore alternative approaches. Let's see after 3,200 tokens of that, what did it come back with? It's come back with, okay, a list of accounts that it should follow for this. And it's basically just going to use requests to try and scrape it directly in here. I'm not sure how successful that would be. Wouldn't surprise me that that would get blocked very quickly. But we can see that the logic in the harness is sound for what it's trying to do in here. Don't forget this is just one shotting it as well. You could set this up to basically keep passing results in. And you could also set it up to actually take the code out and automatically run it and stuff like that. And so, you can see that the other thing too is it actually ends on with each one basically asking, do we want to have some kind of simple Gradio or Streamlit UI, etc. Okay, so I've come over to my chat sort of testing where we've put in just a simple system prompt. It's gone through stuff. It looks like it's found a little bit of a different solution this time. But I want you to see that like that it's now gone and rendered this nicely. You can see we're now in a chat thing where if I basically ask it at the end of that, "Can you build a Gradio interface?" Interestingly, this time the thinking was very quick. So, it understood that it it created the sort of, you know, harness already and that now this was basically just going to structure out a plan to do the rest of the Gradio element of it in here. So, it's not sort of redoing the whole Twitter thing, which is good. We can see that looks like it's made the Gradio app pretty well and pretty quickly in there. So, overall I would say that this is definitely a project that's worth checking out. Try out the other models. I've gone for the 35B N O E, but you could certainly play with this with the smaller 9B model, which is based on the Qwen 3.5 series as well. And let me know in the comments what kind of results you get with this. I don't think this is going to change my daily workhorse for doing coding kind of tasks. I do think if you wanted to do local coding and you don't have access to a big GPU, this is something you could certainly check out. Anyway, thanks to Dell for sponsoring the compute in here. As always, if you found the video useful, please click like and subscribe and I will talk to you in the next video. Bye for now.