NeoHorse-1-4B: The Model That Trains Itself - Run Locally

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: NeoHorse-1-4B: The Model That Trains Itself - Run Locally · Channel: Fahd Mirza · 10:26 · Watch on YouTube

Topics: AI agents, Software engineering with AI, AI content generation, AI models, Local and open-source models

Summary

The video introduces NeoHorse-1-4B, a 4-billion-parameter model built on Qwen 3.5, and focuses on its training method rather than its architecture. The presenter describes how it learns from real AI-router logs and live corrections from a stronger teacher, with improved versions generating further training data. In a local test, the model finds and fixes a serious false-all-clear bug in a glacial flood warning dashboard, and the presenter also reviews benchmarks and a creative-writing test. The model performs impressively at coding, while its writing is fluent but has inconsistencies in characters and setting.

Key points

NeoHorse-1-4B is presented as a 4-billion-parameter model built on Qwen 3.5.
Its distinguishing feature is training on logs from a live routing system that records task types, model choices, and outcomes.
The training process orders real interactions from easy to difficult and includes live correction by a stronger teacher model.
The presenter runs the model locally using vLLM and connects it to a Hermes agent.
In a glacial flood dashboard test, the model identifies and fixes a bug that incorrectly showed zero people in the inundation zone.
After the fix, the dashboard reports more than 41,000 people in the inundation zone and shows overtopped levees.
The presenter says the model compares well with similarly sized models across several benchmark categories.
The creative-writing output is described as structurally strong and fluent, but inconsistent about the country, language, and family relationships.
The presenter’s overall impression is that the model is particularly strong at coding, but less successful at creative writing.

Chapters

Start	Chapter	Summary
00:01	Introducing NeoHorse-1-4B	The presenter introduces the model and explains that its training approach, rather than its older architecture, is the reason for covering it.
01:36	Testing a flood-warning dashboard	The presenter shows a locally running glacial flood dashboard and points out its false-all-clear bug before asking an agent to find and fix it.
03:42	How recursive self-improvement works	The video explains how routing logs and live corrections can be used to train the model, with each improved version producing new interaction data.
05:45	The coding test and bug fix	The model checks the dashboard endpoints and fixes the issue, after which the presenter verifies that the warning data is displayed correctly.
07:20	Benchmarks and creative-writing test	The presenter discusses benchmark results and evaluates a creative piece, praising its fluency while noting inconsistencies in its details.

Quotes

Instead of learning from static examples, this one learned from watching a live routing system decide which AI model should handle which task — 01:04

That’s the worst kind of bug in a system like this a false all clear. — 02:42

So all this self-recursive happening in training not in in inference — 05:13

more than 41,000 people in inundation zone. — 06:46

this is really good at coding no doubt about that. But when it comes to creative writing, I think it is not as good — 09:55

Concepts and entities

NeoHorse-1-4B, Qwen3.5, Token Rhythm, recursive self-improvement (RSI), vLLM, Hermes Agent, GlaciWatch, Fahd Mirza, Master Compute

Full transcript (original language)

00:01 I want to start this video by giving a huge shout out to all the members of the channel. These are the people who keep the lights on. These are the people that I am able to get these models installed locally and show you every day because these people don’t just talk, they help out with their own money. And I’m very very grateful for that. If you also want to help out the channel so that I could keep making these videos which are not

00:32 really cheap then please become a member. It is just $4 to5 a month. That’s about it. Thank you for all the support. The model in question today is this new horse 4 billion a model from token rhythm built on top of quen 3.5. Don’t worry. Yes, it’s an old model. But wait, um, for me to describe why exactly I’m covering this. What makes it different isn’t the architecture. That’s old. It’s how it was trained. Instead of

01:04 learning from static examples, this one learned from watching a live routing system decide which AI model should handle which task, and it studied those decisions to get better. Let’s get it installed and test it out. This is Fad Miraza and I welcome you to the channel. We already have covered that 3.54 billion by the way a lot as you can see on your screen. Recently I also have started this free weekly AI newsletter which you can subscribe at vahadma.substack.com.

01:36 So look at this application which is at the moment running locally on my system. This is Glaci watch a glacial flood early warning dashboard for mountain valley with six towns downstream and I will show you these are the six towns downstream which are under threat. So this is not a toad demo as you can see it is getting the realtime data from my application. Lot of advisory, lot of statistics are happening getting updated

02:09 here in this app which I have already launched with my docker front end, back end and lot of other stuff is running including radius actual database at the back end. Now if you look at this application at the moment it says zero people in inundation zone. If I just simulate the breach and go down, it is going to scroll down here. You see now the breach is happening. The glacier is melting and these towns will soon be

02:42 under threat as you can see from these advisories. But if you go at the top you can see still it shows zero pe people in inund in inundation zone and zero levies over top. This is a bug which is very logical but the implications are huge. That’s the worst kind of bug in a system like this a false all clear. That is where I am going to go and I already have configured this model with Hermes

03:12 agent and I’m going to give it this goal where it need to find and fix every bug. I’m not telling it anything what the bug is, where that is. It just needs to find and do that by using its new technique of self-recursive which I’m going to explain now. This model is running locally as you can see with V LLM and I’ll quickly show you this is how I am serving it. Let me quickly show you the VRAM consumption too which includes KV cache. It is uh

03:42 this is how much it is consuming at the moment. Now while it runs let’s talk bit more around this model. What exactly is difference? I’ll talk about this benchmark in a minute. But before that this is uh the headline idea behind this model. which is recursive self-improvement explainer. Picture a router sitting in front of a bunch of different AI models deciding in real time which one should answer which request. Every time it

04:12 makes the call, it’s also quietly taking notes what kind of task was this, which model got picked and did it actually work out. Normally that’s just operational logging. Neo horse trainers took that exact log and turned it into training data. They fed those real interactions back into the model ordered from easy to hard. So it builds up capability the way you would build up a skill not thrown into the deep end. Then a second pass lets a stronger teacher

04:43 model, you know, watch Neo horse attempt tasks itself live and correct. and it is um in the moment instead of just handing it pre-written answers. That’s the main thing. And here is the loop that you know I believe is the core of this whole thing. Once that improved model goes back into service, it starts generating its own new interaction logs which can train the next version. So all this self-recursive

05:13 happening in training not in in inference by the way maybe that is in uh in the next year or so but that is the whole pitch here the system watching itself work because the fuel for it getting better and better. Let’s go back and this is my GPU card. If you’re looking to rent a GPU by the way you can find the link to master compute in videos description. Let’s check out it is still reasoning and checking it out. Let’s wait for it. Remember just a 4 billion parameter

05:45 model. I couldn’t resist but to show you this thinking. I think the thinking is really stretched for a 4 billion parameter model which is sort of um showing lot of improvement. It is doing lot of recussion. As you can see it checks and then rejects. Now there is lot of thinking. I’m just scrolling through to show you there is lot of recursion happening on each and every endpoint. And then finally I believe that it has checked

06:16 all the end points and returned 200. And there you go. But you can see that during that recursion what is missing is that it has turned to um is this Russian or what? So that is also appearing in its um thinking. Anyway, I’ll just go down to show you still the same. Okay, I believe it is done. Let’s check out the application.

06:46 And I’m just going to do a hard reset on this one. And as soon as I have refreshed it, you can see now it shows us the correct number that more than 41,000 people in inundation zone. And these are the levies which are over topped and everything is in yellow. And you can this see this simulation that which town is in which issue. And there you go. Wow. Red freeboard. So it thought more it recursive loop was

07:20 um stretching a lot but again it has fixed this error. 4 billion parameter model really really punching above its weight after coding. Let’s do a creative writing test. Now read through this prompt because this is very different. This is testing whether the model can turn boring everyday junk, spam, and bills into something emotionally real about lonely loneliness. So, I’m asking you to write an insightful non-fiction piece centered on

07:51 an immigrant living alone using the WhatsApp messages he receives in a single week like junk, e billills, and all the stuff as the core. The piece should explore themes of loneliness and connection in modern society, aiming to be both readable and profound. Let’s run this now. While it runs through it and creates something, a very quick note on the benchmarks. So, it is comparing this Neo horse against five other small models in its

08:22 weight class. 4 billion parameters, roughly the same size. And I think overall the picture is quite interesting and as new horse is on top or near the top across almost every category agentic task coding following instructions it’s not winning by being the biggest model it’s winning because of how it was trained and on the far right that average score across all 10 benchmarks. Um I think that is one number that actually is making sense here because it

08:53 tells you that this isn’t a model that’s good at one narrow thing. It consistently strong across the board and that is I’m also finding it that this is uh quite true. So it is done. Let’s read it out. So I think this is where we have really stretched the model. Um it has done well but there are a lot of issues too. So I think structurally it is quite good

09:24 technically quite fluent no problem and it also has performed aesthetics of literary fiction like fragment white spaces ellipses lot of good stuff is happening here lot of good stuff emojis are there it is well knitted and I’m just scrolling through you can pause and read through it there is um also a mention of China and few others now the issues is that it really never settles on one country. One language or one consistent family relationship like it

09:55 jumps from mother, son, daughter, grandchild all blur into same person. Um which is bit confusing. So I think this is where it has lost the fluency of coin uh after post training in recursive. But what I’m finding that unlike what benchmarks are saying this is really good at coding no doubt about that. But when it comes to creative writing, I think it is not as good as we saw in this 4 billion parameter model from

10:26 Coen. So that’s it. Um, this is what Neoor is. And look, my job is to introduce you to the model, show you how to get it installed, what it can do, what it can’t do, and in my opinion, what it is good at. Um, but of course, I cannot really do your use case in the videos uh because you would know your use case is better. Please understand that um feel free to explore further and then build on top of what you have learned in this video.