Kolibri Has Landed: A Sovereign Open-Weight Model from Germany

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: Kolibri Has Landed: A Sovereign Open-Weight Model from Germany · Channel: Fahd Mirza · 10:35 · Watch on YouTube

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

Summary

The video reviews Kibri, an open-weight model from Germany’s LF Alpha, highlighting its 78 billion total parameters, sparse active parameter count, long context window, and Apache 2 license. The host demonstrates the model generating a playable Blender game and a detailed animated kebab scene using a single HTML file. He also compares Kibri with its unreleased predecessor, discusses a German-language philosophy test, and explains LF Alpha’s iterative training process. Overall, he sees Kibri as a promising step forward, while noting some weaknesses in the English constraint task and philosophical analysis.

Key points

Kibri is an open-weight model by LF Alpha, built mainly for German and English and released under the Apache 2 license.
The model has 78 billion parameters, with about 3.5 billion active per token, and supports up to a million tokens of context.
A Blender Python script generated a playable indoor trampoline park game, with no external assets, in a single attempt.
A separate prompt produced a detailed animated donor kebab scene in one HTML file using plain canvas and JavaScript.
Compared with the unreleased Kibri Origin, Kibri grew from about 30 billion to 78 billion parameters while its active parameter count changed only slightly.
LF Alpha reports that training data increased from 7.5 trillion to 20 trillion tokens, and training context length grew from 64,000 to 262,000 tokens.
The German philosophy test suggested strengths in German writing and following constraints, but the host found the English response somewhat vague and the final analysis weaker.
The training diagram describes repeated small development-model experiments and evaluations before the costly final training runs.

Chapters

Start	Chapter	Summary
00:01	Introducing Kibri	The host introduces LF Alpha’s model, its specifications, license, context length, and local VRAM requirements.
01:06	Game and graphics demonstrations	The video shows a Blender-generated trampoline park game and a single-file animated kebab scene.
04:14	Kibri’s development and model comparison	The host describes Kibri Origin and summarizes the reported changes in Kibri’s size, training data, context, and architecture.
05:53	German philosophy challenge	The host evaluates Kibri’s German and English responses to a multi-part prompt about Hegel, noting strengths and shortcomings.
09:01	Training process and closing thoughts	The video explains LF Alpha’s iterative training workflow and concludes with the host’s positive assessment of Kibri.

Quotes

Now, Kibri is a new openweight model from this German company built mainly for German and English. — 00:01

it was able to build it in just one shot. No external assets at all running all locally. — 02:41

That means it got much bigger without getting much more expensive to run. — 04:46

it seems that they are bit vague and flat and the English loses the owl entirely. — 08:27

In short, the real work is hundreds of small experimentations and the big run is just the last step. — 10:03

Concepts and entities

Kolibri, Kolibri Origin, LF Alpha, Godot, Blender, Apache License 2.0, Hegel, HTML Canvas

Full transcript (original language)

00:01 This is Kibri, a German word for hummingbird. The company behind this model is LF Alpha, which we have been covering from day one. As you can see, they released their first model more than 2 years ago. They are back in the game. Now, Kibri is a new openweight model from this German company built mainly for German and English. It has got 78 billion parameters in total but only about 3.5 billions are active for

00:33 each token. It can handle up to a million tokens of context and it is released under the Apache 2 license. At the moment it is taking around 140 gig of VRAM on my local system running two GPUs courtesy mass compute. If you’re looking to rent a GPU on very good price, you can find the link to mass computing videos description. This is Fahad Midsa and I welcome you to the channel. Right away, let me show you a result which was done by this model. So,

01:06 this test is not about making a pretty game. It is about whether a model can work like a a patient developer. As you can see here, I ask it to write a Blender Python script and you can see it is running. I’m pressing my space bar and it is running pretty nicely. Now this is a Blender Python application with Go dot and this has built a complete indoor trampoline park, the warehouse, the

01:38 trampoline grid, a foam pit, a dodgeball court, a slam dunk zone and a ninja warrior course. This is what I asked it to do. Took around 40 to 60 minutes and then it has to export everything which it has done. Build a playable Go.4 4 game with a third person player. Let me press space bar again. Let me press E and it should it takes bit of a time first time. So let’s wait for it. Okay, I think I messed it up. Let me restart

02:10 it. There you go. So this is E and this is Q. You see it is spinning. It is spinning. Spinning. Spinning. It is pretty good. Now you can also press F to do a spin. and again space bar and I’m continuously pressing the space bar here. It is also showing me all the physics which is involved different angles and now I’m moving the mouse. Of course there are lot of rough rough

02:41 edges but you see that it was able to build it in just one shot. No external assets at all running all locally. Let me show you another result which this model has done. Look at this one. This one is a different kind of challenge which I gave it to this model. So a very long very detailed spec and asked it for a single HTML file with no libraries, no images and no outside help. Inside it there should be a cinematic donor kebab

03:11 spinning in front of gas burner with blue flames, glowing heat, smoke, embers, grease, sparks, shimmer, shimmering heat, distortion and stainless steel reflections all drawn with nothing but plain canvas and JavaScript. It tests whether the model can follow dozens of instructions at once and turn them into a real graphics code without forgetting half the list along the way. And you can see that how many items on that list it has actually

03:42 delivered almost all of them. My favorite part is these flames on the right hand side panel. I think they look real good, realistic and works like a charm. Um, you can also increase the speed here on the top left. Slow and then normal and then fast. Let me increase the heat. This is a low heat. And you can see the color of it. This is very high. Again, low, medium,

04:14 and high. We can pause it and then we can resume it. All happening in one file with this Cibri model. Now the story doesn’t end here. They have shared lot of documentation, technical post and lot of other things. Out of all of those, this is my favorite part because it tells a story and I love stories. Collibri has an older sibling called Cibri Origin which was never released to the public. LF Alpha says

04:46 they built it first to test their training pipeline and then use the same pipeline to build Calibri. About 3 months later, look at what changed in that short time. The model grew from about 30 billion parameters to 78 billion. Yet the number of active parameters barely moved from 3.27 billion to 3.46 billion. That means it got much bigger without getting much more expensive to run. The training data went from 7.5 trillion tokens to 20

05:19 trillion and the context length it was trained on jumped from 64,000 to 262,000 tokens. Experts inside the model tripled from 128 to 384. Its attention also changed so that most layers only look at nearby text and just a few look at everything which is a big reason long context stays affordable. It also learned to think at four different levels instead of one and its knowledge now goes up to June 2026 instead of 2024

05:53 which is very impressive. These are LF Alpha’s own number by the way. But if they hold up and I think uh to some extent they do in the test which we have just seen. This is a lot of progress in one quarter. Let me show you a multi- lingual test in German. The results. So this is the prompt which I used. This is a five-part philosophy challenge built around Hegel’s famous owl of Manurva line. I have asked the model to

06:23 translate it two ways. Explain its meaning. Write dense Hegelian German. Squeeze it into strict English. Then name what got lost. It is very hard because it stacks translation, deep philosophical understanding, creative German writing and hard constraints in one go and one weak step shows up in the next. It really matters for Kibri because German and English are the whole point of this model. Now this is the

06:53 fairest test of its core promise in my opinion and this is the result. I’ll be upfront. I really cannot tell you this answer um you know with full certaintity because I really don’t speak German. I have uh I’m not an expert in you know Hegel at all. I’m not even sure if I’m pronouncing his name right. But if you are German or if you know about Hegel, please please guide us, advise us as I’m not an expert at all. But you know with

07:25 different judge models what I can tell it that it opens well with a literal and poetic translation that really do differ and a clean explanation of why the Hegel section is solid in my opinion. Owl as philosophy, dusk as an epoch ending and a clear breakdown of why um of bang means cancelling, preserving and lifting up at once. My apologies if I’m not pronouncing German right. The German paragraph is the showpiece as per the

07:57 judge model. It is one enormous nested sentence uses of correctly and sounds properly hegelian. Though at roughly 100 words, it runs a bit long, I would say. It does contain um some other intelligence something which is fine since the ban was only for English version and then comes the trap in task four which I was very interested I’ll just go there. There you go. The three sentences are short and avoid every band word which I

08:27 asked. So it passes the rules. Uh but the thing is that it seems that they are bit vague and flat and the English loses the owl entirely. So the idea itself gets thin. Task five is the weakest part. The three lost nuances are reasonable. But the third one just repeats the off point from the task two. So I think as per the judge model’s reasoning, it is quite strong on German and on falling constraints a little shallow where the real insight was needed. Also I think overall the result

09:01 is quite interesting. One more thing which I really want to show you is you know around its models training where they say that model training as a code and look at this diagram. So this diagram primarily shows how this model got made and the surprise is that nobody trains it just once. On the left is pre-training. The team picks some data, a recipe and an architecture, then trains a small development model first, checks how it does and loops back to

09:33 optimize things. They repeat that loop in you know any number of times running many cheap experiments and only when the setup looks right do they commit to the big one. Expensive run that produces a base model on the right post training works the same way. Now the ingredients are data, a recipe and rewards which teach the model how to behave. Again there is small development model and evaluation and a loop repeated m number of times before the final training run

10:03 turns the base model into the finished one. In short, the real work is hundreds of small experimentations and the big run is just the last step. And yes, don’t worry. I’m going to still tell you to subscribe to this newsletter because still um the number of subscriptions are not as good as I would like. So, please subscribe to this free weekly AI newsletter at vajba.substack.com. If you want to help out the channel, please become a member. The link is at the homepage or in the first pin comment. Look, Kibri is less um one big

10:35 trending rent than a small bets. I really like it and I think um a real good step up for this LF Alpha team. That’s it. Let me know your thoughts. Thank you for all the support.