Run Ornith-1.5 122B Locally: Stronger, Faster, Leaner

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: Run Ornith-1.5 122B Locally: Stronger, Faster, Leaner · Channel: Fahd Mirza · 14:01 · Watch on YouTube

Topics: Software engineering with AI, Local and open-source models, Multimodal models

Summary

The video tests Ornith-1.5 122B locally with llama.cpp across coding, image understanding, creative coding, scientific reasoning, screenshot interpretation, and translation. Its coding agent finds and fixes a faulty inheritance rule in a full-stack application, and the presenter reports strong results in several other tests. The model also produces a detailed curry recipe and an animated kangaroo tattoo in a standalone HTML file. The presenter finds it a capable open model for everyday coding and agentic work, while noting weaknesses in some low-resource language translations.

Key points

The presenter runs Ornith-1.5 122B locally with llama.cpp and uses Harvey’s agent to test whether it can fix a bug without hand-holding.
The model corrects an inheritance bug: Thomas, who died before Margaret, should receive nothing, while Walter receives his share.
The presenter describes the model as an open-weights mixture-of-experts model released under the MIT license, with a context window of over 260,000 tokens.
In an image test, the model identifies a curry as Mughlai and writes a recipe with quantities, steps, and tips.
It creates a standalone HTML animation of a kangaroo tattoo being drawn, with no external dependencies.
In a visual scientific-reasoning test, it identifies A and D as receiving an electric shock and explains its reasoning.
It interprets a WhatsApp screenshot and proposes humorous, context-aware replies to the wife and the boss.
In a translation test covering 81 languages, the presenter praises several major-language translations but finds some low-resource and regional-language outputs weak.
The presenter concludes that the model is a good all-rounder for coding and agentic work, though not perfect.

Chapters

Start	Chapter	Summary

Quotes

the model says it has fixed the bug — 04:11

without any external dependencies — 08:47

It has read the screenshot very correctly. — 10:54

I think a good all-rounder model. — 13:29

Concepts and entities

Ornith-1.5 122B, llama.cpp, Legacy Vault, Atomic Chat, Mixture of Experts, quantização, Fahd Mirza

Full transcript (original language)

00:01 Ornith 1.5 122 billion parameter model is here and I’m going to test it straight on this Legacy Vault application. This model is said to be faster and leaner from its predecessor and that is exactly what I’m going to test out in addition to various benchmarks. Legacy Vault is a tool law firms use after someone dies. It reads the will, the assets, the debts, and the family, then works out

00:34 who inherits how much. The rule in this will and here must outlive the person who died by 30 days or they get nothing as you can see on your screen. But this app has this rule backwards. Look at the screen. Thomas died 132 days before Margaret. He’s dead, but the app pays him Australian dollar 1.95 million. And then, you know, the bug just spews through and leaks through

01:04 this application and that is what I’m going to fix with this new model. This is the application, a full stack application running with Docker with a lot of components on my system. And this is where I’m already serving this Ornith 1.5 model with llama.cpp with this command and this is how much VRAM it is consuming. At the moment, I’m using this GGML contest version from Atomic Chat and you can also select any

01:35 quant of your choice as per your own GPU VRAM. Let me quickly show you the VRAM too. So, you can see that this is the VRAM consumption at the moment. The reason why it is so high is because of my context length. Of course, you can also reduce it easily. I already have installed Harvey’s agent and that is where I’m going to give this model a goal without any hand-holding to see if it is able to find that bug and fix it. Also, recently I have started this free

02:08 weekly AI newsletter, which you can subscribe at farhatmirza.substack.com. And if you’re looking to rent a GPU on very good price, you can find the link to my compute in videos description. While the model works through that prompt, let’s talk a bit more around this model. So, the best thing about Oracles that it is open weights model released under the MIT license. So, you can download it, run it, and build on it freely. It’s a mixture of

02:38 expert model with about 122 billion parameters. It reasons before it answers. It reads images as well and text and handles a context window of over 260,000 tokens, which is bit on a lesser side, but that is fine. The pitch here is three words: stronger, faster, and leaner. They say that it finishes task quicker and writes shorter outputs without wasting tokens. And it follows instructions better and breaks big jobs

03:09 into step more cleanly. It is built for real work like coding, agents, and tool use, not just chat. It is still working though on our prompt, so we will wait for it. Let’s have a quick look at the benchmarks. The interesting part is who is standing next to it. It beats model of its own size like 1.3.8 flash next and Nvidia’s NeMo Tron 3.5 super on most of coding and agentic test. On SweepBench Pro, which uses real

03:40 software issues, it’s far ahead of everything else on the chart. Then it goes toe-to-toe with DeepSeek before flash, which is more than twice its size, and with Beam, which is about four times bigger. It doesn’t win everything, but a model this size keeping that company is quite impressive. So, anyway, instead of believing on these numbers, we will check out this prompt to see how it goes. So, you see it already has confirmed few things, but I will let it run and then

04:11 we will check out the end result. After thinking long and hard, the model says it has fixed the bug and it it has also caught some of the false alarms, which is quite good. Verification also looks pretty good. So, it has taken its time, but not too long. I think it’s a balance in reasoning plus the result. But anyway, let me now go back to that application and do the hard refresh. And the bug is definitely fixed, as you can see. Thomas who died before Margaret

04:42 now gets nothing and Walter who outlived now gets his 519,000 Australian dollar share. So, the model found the one swipe line on its own, rebuilt the app, and the ledger still balances. So, very good result. And now let’s do a vision test. I’m giving this model a photo of a curry and asking it to name the dish, say where it’s from, and write a full recipe. So, primarily, I’m testing whether it truly sees the

05:12 image and can can turn that into useful accurate detail. Let’s run it in llama.cpp server GUI. And we will check out what exactly the response is because it also includes some of the creative writing. The model has given me the response, but even before I show you that response, I really want to spend a minute on this reasoning. So, if you go through it and I’ll scroll down, it didn’t just guess. It weighed Thai red curry, Caribbean curry, and a

05:43 general Indian curry, spotted an odd sausage-like slice in the image which you can see again. And then checked itself before committing to one dish as you can see from this reasoning. The thinking was thorough but it went back and forth on a lot on one photo. So, which is good because it wanted the answer to be correct. So, but I think it is not as lean in this high thinking. You can always reduce the thinking.

06:13 So, but it is fast for sure. In terms of answer it is quite strong too. It named the dish as Mughlai a North Indian chicken and potato curry and pointed to exactly what it saw. Some chunks, chicken pieces, red chili oil and few other things. And then it has written a full recipe for four with quantities, 10 clear steps and tips which is what I asked for. So, really really good stuff. You see and then some tips too. So, fairly good answer I would say and

06:43 if I really compare this new version to the previous version from Ornet, this one is way way better and really one thing I want to say directly to the Ornet team that I have been covering their model from day one when no one even knew about this Ornet. And I am looking at some of the you know, they have been marketing it on X and few other platforms where they gave people early access but they didn’t even ask us who have been their day zero supporter. So, would

07:14 be good that if they remember us for the next video. Anyway, so our next model release. But I think still, you know, regardless of that I think this model is much much improved version. And now let’s do a creative coding test. I’m asking it to build a single HTML file that animates a kangaroo tattoo being drawn on a thigh. So, I’m testing visual design, animation code and whether it gets the complete result in which is quite complex too in one go or

07:46 not. The model has created the file. Let me open it in front of you. I haven’t seen it yet, so let’s check this out together. There you go. So, simulation is happening. It is creating some sort of tattoo here. And it is it has created the outline. Now trying to shade it. Creating some more art there. And you see that looks pretty good so far. Oh my goodness, look at this

08:17 shading now. So, after creating the silhouette, you can see that it is now creating all the angular figure. Very very well done. Very well done. And now creating some of the other stuff. I think this is pretty good. This is pretty good. Yes, I mean uh this is not an image generation software, so we can imagine that it is not going to generate a that sort of quality, but still it is

08:47 quite a world-class output in terms of just creating it in single HTML file without any external dependencies. So, this illustration of Australian traditional style with solid black fills and a limited palette is quite good. The stars are also scattered. Pretty good. I’m just running it again. You can see that this looks so good. So, so good. I’ll increase the speed. There you go. Let’s put it to max.

09:18 Yep. So, this is higher speed and now it is creating it with you know, much faster speed. Pretty good. Pretty good. Sorry, I am just enjoying this tattoo. I think this looks pretty good. Yes. There are defects, of course, but as I said, it is just um a typical HTML file. Nothing fancy is happening here. Let’s do another test. Next up, let’s do a check on scientific

09:50 reasoning. I’m just going to give it this prompt where people are holding on to this different wires, and I’m asking it who will get an electric shock. So, without giving any text, let’s give it this prompt. Let’s see if it is able to infer that question from that image. And quite swiftly, model has given me the correct answer, A and D. So, it also explained that a shock needs a complete path through the body, and that D is shocked even on a stool because

10:21 um he touches both the wires, and then this is again a table of summary it has generated. So, a real good answer there. Okay, now let’s check out this test, totally hypothet ical, but I think the model has given the best answer so far. Look at the test. This is a WhatsApp chat where a boss and 11:00 p.m. deploy and a suspicious wife all collide. So, we are testing whether it reads the screenshot correctly, and then finds a smart funny way out of an almost

10:54 impossible spot. I wish I could show you the whole reasoning which it has done because it is so so good, so fun to read, but I will skip that for now. Let’s check out the response. It has read the screenshot very correctly. It worked out who said what, and that tonight means the 11:00 p.m. deploy, not a fling, and that the wife only saw the first two messages. So, very very nuanced model. It solved the problem in a smarter order. Look at

11:24 this message to the wife. Look at this. It’s a deploy, not a dossier. How good is that? So, and then, you know, it replied to wife first, then boss, and then the deploy and gave a clear reason. The boss reply approves his leave or her leave. It is a right direction and it asked permission to reassure the wife. And then, you know, one line plan for tonight. Be home by this and then kitchen light on so your wife watches you. Be gloriously boringly innocent.

11:55 Very very good writing skills. Very happy with the model. And now, let’s do a multilingual test. So, this is a situation where a man is religiously mowing his lawn and so much so that even the grass is gone because he couldn’t really take his eyes off as the neighbor won’t stop jumping on that trampoline as you can already see. So, what man is trying to do here, it is asking the, you know, the lady to have

12:27 mercy on me and my lawn. So, and this is what this model needs to translate in all these 80 languages. So, let’s see how the model goes about it. Please check your language and let me know in the comments when it is done if that is correct or not. Okay, here is the answer. First thing which jumps out is that it didn’t really follow the instruction to the hilt because it has changed the sentence, but anyway, um it finished all 81 languages native

12:58 scripts one line each and the major languages look really right to me like Spanish, French, German, Korean, Japanese, Hebrew, Arabic, all look okay to me, but let me know your thoughts. But then, you know, as usual, like various other models, the rare ones or the low-resource languages are weak. Like Thai, like Hawaiian, Zulu, all look invented to me. It doesn’t really look good. Even the regional Indian or Pakistani languages doesn’t look good.

13:29 Southeastern, um, languages also, I think, fall off. And I was expecting better from this model in terms of humor. But I think, um, it has done an okay job, if not, uh, you know, world-class. We will take it as a good one. So, look, I think a good all-rounder model. It was quite good in almost every test. Capable open model for coding and agentic work, if not the perfect one. But I think that is what we need. We just need, uh, daily worker. For that, I

14:01 think this is a good one. Let me know your thoughts in the comments. Thank you for all the support.