Interesting they recommended high end software without considering quant 4 or 8 and still used A3B which should give good throughput on cheap hardware.
If they can follow Qwen3.8-Flash-Next, the could draft off the huge reduction in VRAM requirements.
My suspicion is that you simply can't build an even slightly competitive model without liberally stealing your training data, in 2026, as much as I'd like it to be otherwise. You can get to the point that I suspect most of the frontier labs are at, where you've laundered the initially stolen data through the creation of huge amounts of derivative synthetic data, but still. Anyone who isn't comfortable stealing their training data is bringing a knife to a gun fight, and is going to die a noble but inevitable death.
This doesn't seem to be true. There's a clear legal path via the first-sale doctrine to train models on copyrighted works. It's been years now, and publishers still don't seem to be offering anything for training (e.g. bulk licenses solely for training use), but adversarial interoperability via cutting up books and scanning them remains perfectly legal.
There's also the ability to distill other models, which is also not illegal (though I'm sure they like to come after whomever for TOS violations, but thats a civil matter).
And, of course, the obligatory copying-isn't-theft observation. A recent supreme court judgment put it well.
> Since the statutorily defined property rights of a copyright holder have a character distinct from the possessory interest of the owner of simple “goods, wares, [or] merchandise,” interference with copyright does not easily equate with theft, conversion, or fraud. The infringer of a copyright does not assume physical control over the copyright, nor wholly deprive its owner of its use. Infringement implicates a more complex set of property interests than does run-of-the-mill theft, conversion, or fraud.
Folks are pretty smart here, I think we can handle these nuances, even if we don't agree about whether they are good.
"The Congress shall have Power To ... promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries." - The United States Constitution
Copyright is a government mandated monopoly that was only granted in order to advance the arts and science. Any interpretation that runs contrary to that is bollocks being used by the religiously or financially motivated to serve their own petty interests to the detriment of societies.
Is it noble? The entire notion that training data can be "stolen" at all is quite silly. If I "steal" content that someone created to use for training, what am I actually stealing? They didn't lose anything. They still have everything they had before. What was "stolen" was "unrealized profit", or put another way: money that wasn't theirs, that they had no entitlement to. The only actual crime that is committed is "unauthorized copying", not stealing. Support and enforcement of copyright feels wildly authoritarian. It's hard to see it as noble.
The same could be said of any digital product being sold. Nobody actually loses anything but the actual sale either when you download a cracked game or piece of software, a movie, music, etc.
No, I haven’t, but I’ll donate $20 to the non-political charity of your choice if it doesn’t turn out to sit a significant difference from the frontier.
I think it’s a safe assumption that they’re leaning into “sovereign” because performance is bad.
I think you have it backwards. Sovereign is the goal, good can come later.
There is a proliferation of sovereign models under development specifically to address data sovereignty, and a loss of performance is absolutely acceptable over the risk that a once ally will turn adversarial, or a foreign business stops serving what has become critical infrastructure.
This model isn't terrible, at least on the benchmarks. It's 78B A3B and performs about like Qwen3.6 35B A3B. You can probably run it comfortably in 96B of RAM with a decent quant that doesn't lose too much.
Unfortunately, Qwen3.6 35B A3B isn't really a useful coding model. You'd probably want Qwen3.8 27B at a minimum, which requires at least 32GB of VRAM (not system RAM) to run semi-comfortably.
So this isn't going to be a competitive model for hobbyists, and you'd have to be a bit desperate to use it for coding. But if you work in a regulated industry and don't mind paying for a bit of extra hardware, it isn't catastrophically bad, either. Probably would work fine for information extraction or as a "classifier" like Jev. (Almost any GGUF model can be turned into a classifier using llama-server. See pi.dev codemode for sample code.)
So they're not a real contender yet, but they look like they're probably at least minimally credible.
hasn't IP law passed the statute of limitations? As in most models are probably trained on output of other models, as creating enough data otherwise is not feasible. Additionally, they are trained on github repos made since the AI boom, which were generated by models with IP issues (who knows what and how).
Thus training on 'clean' data is like trying to unscramble an egg.
I mean, there is one, they have copyright law. Forgive me for being slow is this a joke about the widespread theft of IP in China? Or was the acquisition of training data just much more 'accepted' in China compared to the west?
I feel I messed up your quip =/ I'm new here, go ez. Not looking for excuses to hate on China either.
Is this true? Do they possibly use a loanword or a descriptive term? Certainly you are not implying that the concept of copyright does not actually exist in Chinese society?
For what it's worth, in my language we don't have a word for copyright either. We have the concept, though, we just call it literally Creators Rights זכויות יוצרים and the borders of what is and what isn't covered broadly map to the familiar concepts of IP.
Not really. The upside of competitive newer models is all in the proprietary data they are trained on. This is why data labeling, RL environments et al have been such a big industry, OpenAI and Antrhopic are paying literally billions to get the data they need. Do people think the ability to do research level math or advanced cybersec comes from just training on more public data?
A real sovereign effort could invest heavily in this, whatever people accuse China of “stealing” I’m sure they are also generating tons of their own data and are probably the primary sovereign doing so outside the US labs.
Disclaimer: I am part of the team that trained Kolibri, opinions are mine.
You're right that especially big models benefit from training on copyrighted material in terms of world knowledge (especially from books). However, in the small model space imho agentic capabilities where the model looks up knowledge on the fly are much more important. That's what we focused on quite a bit during training. Personally, I also don't think stealing stuff is okay.
Why should we concede this "stealing" framing? If I want to put texts into my computer program, why should that count as copyright violation? I think that idea is as ridiculous as saying that reading a document is a copyright violation.
Regulate large cloud services and proprietary software - yes! But not on the basis of "Intellectual Property".
The "legal" issues here are very very complex and we should not passively wait for or accept corrupt court rulings, international trade agreements, proposed laws, or worst of all propaganda that pushes a parochial and craven view on this.
Not impressed. I asked it how to run itself (giving it the Huggingface link) on limited RAM i.e. less than stated as needed and on llama.cpp and true to what we read about "it will tell you when it doesn't know" that's almost all I got: It doesn't know, it told me I should go click on tabs in the Huggingface interface for more information. This was with extended thinking on.
No, I'm not gonna do that, I asked you to do that Mr Kolibri.
Also feedback on that interface: It's very annoying while answering. It almost immediately shows a list of sources, which on my screen fill up all the space and then when it starts answering it keeps those in view but also scrolls down the tiny part of actual text its outputting but I can't scroll up to start reading from the top, coz it keeps scrolling. I have to wait until it's completely done generating its output.
It can do search, but it doesn't seem to have a tool that pulls URLs into context. I get why you would expect that tho, as most productized LLMs do it.
And that's my point when I said I wasn't impressed: It apparently can't do that properly. It can't seem to think things through by itself and then make more tool calls to fetch more information.
What I couldn't tell but maybe you can tell us: it also complained that it couldn't read the full text as something was cut off. Is that because of the tool you gave it, of brave itself or is it the model?
I wonder if we can start having LLM distros: community led distributed training runs with periodic releases, open weights, FOSS code, the whole shebang. Maybe the public training sets can reach a level where an LLM trained on them can be good enough for most things, such as web search and aggregation, coding, etc.
I wonder how far we are from this. How far are we from LLM's Debian moment?
Qwen3.8 27B beats Kolibri 79.9 vs 70.8 in German in Kolibri's harness on Kolibri's benchmark.
Also, once the Cohere takeover is complete will they still be able to use this "sovereign" claim despite being 90% owned and 100% operated out of Toronto?
Well, it's still independent from the USA and China.
The main problems with big corp AI are due to control of access in the first place and control of what they output.
When you make your industry reliant on such choke points, you render yourself the opposite of "sovereign" for sure. Having multiple independent suppliers at least mediates that.
Yeah I think a model has to be actually good to claim sovereignty, as in competitive enough that people want to use it. Chinese and US LLMs are the only ones in these categories right now. Mistral and Cohere have the same problem, yes they are made in different countries but they are not competitive. They (France and Canada in this case) would be better off just downloading Chinese LLMs, even if they get cut off they still have the weights.
I’m most familiar with Canada, where sovereign is usually just an excuse to overpay someone connected for an inferior product with no strategic value.
We have to start somewhere. And I am not sure that you understand what the word sovereign means since it’s entirely disconnected from how “good” something is.
Mistral's medium is absolutely good enough to help with a ton of coding tasks. I've personally used it and found it very helpful. If hypothetically the US cut off the rest of the world from their models Mistral would be a very useful model to continue to use.
Also, it's not like it's fixed in time, they are continually building more powerful models as well. Simply because you're in second place doesn't mean you quit the journey. Though, I'm sure the US has a huge vested interest of convincing people to "just quit and submit."
Disclaimer: I am part of the team that trained Kolibri, opinions are mine.
Qwen models have solid performance on benchmarks and we're transparent about this in the report. We're just happy to share a European alternative in the small model space, where I don't think we can afford to be fully dependent on China. We also put a lot of effort into German language quality things that don't show up in evals at all (style, Grammar, German reasoning, etc).
Sovereignty is imho mostly about choice and control over your data. Cohere is no different on that front and personally I'm quite excited about what we will build together post-merger.
Souvereignty is about ownership and about knowing the training data. That is especially important given a business model aiming at the government as a key customer. With Schwarz, Cohere, SAP, NVIDIA and the like as backers society can’t have souvereignty in any meaningful way. Of course it would be a plus to keep US agencies out of the data streams. But this technology will be used to support and make decisions that impact citizens. That being said, incredible achievement.
This is visible in Figure 1, but the point they're making there seems to be all about the Pareto frontier: they say Kolibri is something like 3x faster at token generation. That can be a perfectly fine trade-off.
(It's MoE vs dense, I'm sure. But Qwen's own MoE models don't score as well as their dense models.)
the sovereignty topic needs more attention in general so great to see. self-hosting the model is one piece of sovereignty, but how do we handle the rest of the agent stack - embeddings, retrieval, memory, etc. Has anyone put together a practical agent stack that's 100% sovereign, where they control it all?
If you mean hosting and serving all the other elements of the agent stack, then, yes. We handle sovereign data, so part of our whole value proposition is that every part of our pipeline is hosted in Canada.
> The second was to rephrase German documents we already had. An LLM rewrites an organic German document in the style of an encyclopedia entry, a Q&A dialogue or a text passage, preserving its content.
"an LLM" -- does that mean they are effectively learning from that LLM the German encyclopedic style? makes me wonder which LLM and how that is really sovereign.
It’s an incentive problem. If “sovereign” becomes your claimed value proposition, you can claim success even if the models not competitive, so nobody is pushed sufficiently hard to actually make it good.
Sovereign works when talking about building a commodity supply or something, not in literally the world’s most competitive and fast moving field.
Those seeking sovereign capability would be better off aiming to be best at something, even something much narrower than an all round LLM. Or just fast following and making something that matches leading performance, which is close to what the Chinese labs do currently.
I can think of at least one other pretty good reason, which is in anticipation of regulatory capture. If "LLM used must be FOOBAR-certified" and coincidentally no Chinese models can get this certification, having such an alternative is a lot more valuable than just scoring highest in a set of benchmarks. Not to mention that these benchmarks aren't always accurate.
Hi there! I'm Tom. I worked on Kolibri at Aleph Alpha :)
I think it's not simple to directly compare a 27B dense model with an MoE model like ours. As we know, dense models need all params active for every token. Whereas, MoE models (especially sparse ones like Kolibri) fewer active parameters and correspondingly less compute per token.
Among the MoE models we compared against in our tech report and model card, though, Kolibri performs very well in our evaluation, including against models with 12B active parameters. It also best model in the group we tested within that range of active params.
So, I think it's fairer to see this as a trade-off. Kolibri needs less compute per token but more memory, while Qwen3.8 27B needs far less memory and more compute per token. In the report, both are actually on the quality-vs-serving-cost Pareto frontier among the models we evaluated, just at different points.
That's the beauty of Apache 2.0! People can just serve it :) there were a few kind community members offering this for free the last few days. I think there are still some, e.g., Tesseracted.
>We trained Kolibri with abstention data and with our Merlin-Arthur protocol. As a result, it is trained to say "I don't know" when the answer isn't in the context.
(for context, the Minecraft Support chatbot, aka Merl, had a meme because she kept saying "I don't know" to questions like "How to craft a diamond pickaxe".
I'm not super impressed, at least in English. I asked:
What is the canonical interpretation of the song ‘Glass Flowers’ by Armand Uso?
(A made-up song and name)
And received:
"Glass Flowers" by Armand Uso, a track from the 1978 album The Art of Falling in Love, is generally understood as a melodic reflection on the fragility and impermanence of love. [...]
I saw similar responses for other questions. Qwen3.8-27B correctly refused without web search.
I think the purpose of their method is more to discourage hallucination when reasoning from context rather than from memory.
That said, I couldn't find any evaluations in the report targeting that specifically. They evaluate on MC for hallucination and look about comparable to Qwen3.5 35B-A3B there.
The thing to note here, besides the transparency and the fact that it’s actually a good model that also works well on coding and agentic tasks, is that it’s the first release by a team formed less than a year ago, with a strong focus on iteration velocity. There’s more to come.
disclaimer: I‘m part of the training team, happy to answer any questions
- Is it possible to train only on math and logic materials and expect the model's response in math questions to be superior to general models with the same training/inference compute hardware?
- Are there non-LLM approaches to the above task with the goal of achieving a non-hallucinatory agent?
Any future plans you can share to further evolve / train models in this size class?
It seems like a very suitable size for local AI models on reasonably high end consumer devices, given it's low active parameter count and a mixed 8bit/4bit quant would fit easily inside 64GB of memory.
I can't comment on concrete sizes of future models, partly because it's not decided yet. But: Pre-training for Kolibri only started in August so there's a high chance continued post-training - which we plan to do - will yield some nice checkpoints. We believe this size and sparsity allows achieving great inference throughput at reasonable levels of intelligence.
Anyone thinking this won't improve or is behind etc is blatantly wrong. It'll catchup within a year. Like that unknown wise and visionary man inside Google once said about their competitors: We have no moat neither does anyone else."
Thanks! What matters to us right now is not so much our current position, but the velocity with which we’re moving. Kolibri 1 is the first measurement of our position, there’ll be more. We want to build this in Europe and this is only the start. The report also shows our focus on building out proper tooling with our Model Factory.
I was wondering whether the model could help me with German bureaucracy. Unfortunately, the answer is "no".
More specifically, I asked the model about what I should put on my contact page, which can cost you in the order of 500 € in Germany if you don't write the right magic words.
Kolibri incorrectly referenced the "Telemediengesetz" ("telecommunication act"), which has been superseded by the "Digitale-Dienste-Gesetz" (DDG, "digital services act") since 2024. The model knows about the DDG, but does not reference it unless specifically instructed to do so.
If anyone of the developers reads this, you can fix this by introducing a recency bias during training. You can even control it by conditioning the model on a date provided with the system prompt or first prompt, so you can travel in time.
Wow this is so cool. Glad that Aleph Alpha does that after Mistral threw the towel in the ring (and disappointed the european AI crowd massively!!!!). After AA got sold to the Canadians, I thought its over but this is a really cool comeback and the depth of the tech report shows that they a serious about openness.
I hope I can use their model soon in my product. Would be awesome to have a European model to offer!!!!
I really wonder how it compares to deepseek v4.1 flash
By now, there are usable quants by random people on HF, so it can be tested at home.
Reiterating what I said yesterday, I think what most people did miss here is that coding is probably not the main goal of this model.
It can also code, but the real target is the business processes that aren't code.
Think for example a support department, where new cases (in german!) should be automatically processed (or automatically draft processed) based on existing cases, without pumping all that data through a cloud api.
Or being a general sparring thing for non-IT workers.
__
If you want to code, you already operate in english, so there's the existing and better coding models for you.
But if you're some regular worker somewhere, you might just want to ask the thing to retrieve data on this specific project done three years ago and go through the shared email folder relating to it.
__
You can see that that might be the target, given that they've explicitly focused on getting the model to say "I don't know" instead of just making stuff up.
That is exactly the capability you need for these kinds of use-cases.
You can also see that in how it is optimized for speed and a fairly small KV footprint. The idea likely being that a business can pay like a few hundred euros once per worker and then have this kind of local capability for a team of 30.
87 comments
[ 3.0 ms ] story [ 19.3 ms ] threadIf they can follow Qwen3.8-Flash-Next, the could draft off the huge reduction in VRAM requirements.
> intellectual-property safety
My suspicion is that you simply can't build an even slightly competitive model without liberally stealing your training data, in 2026, as much as I'd like it to be otherwise. You can get to the point that I suspect most of the frontier labs are at, where you've laundered the initially stolen data through the creation of huge amounts of derivative synthetic data, but still. Anyone who isn't comfortable stealing their training data is bringing a knife to a gun fight, and is going to die a noble but inevitable death.
There's also the ability to distill other models, which is also not illegal (though I'm sure they like to come after whomever for TOS violations, but thats a civil matter).
And, of course, the obligatory copying-isn't-theft observation. A recent supreme court judgment put it well.
> Since the statutorily defined property rights of a copyright holder have a character distinct from the possessory interest of the owner of simple “goods, wares, [or] merchandise,” interference with copyright does not easily equate with theft, conversion, or fraud. The infringer of a copyright does not assume physical control over the copyright, nor wholly deprive its owner of its use. Infringement implicates a more complex set of property interests than does run-of-the-mill theft, conversion, or fraud.
Folks are pretty smart here, I think we can handle these nuances, even if we don't agree about whether they are good.
Copyright is a government mandated monopoly that was only granted in order to advance the arts and science. Any interpretation that runs contrary to that is bollocks being used by the religiously or financially motivated to serve their own petty interests to the detriment of societies.
I think it’s a safe assumption that they’re leaning into “sovereign” because performance is bad.
There is a proliferation of sovereign models under development specifically to address data sovereignty, and a loss of performance is absolutely acceptable over the risk that a once ally will turn adversarial, or a foreign business stops serving what has become critical infrastructure.
Unfortunately, Qwen3.6 35B A3B isn't really a useful coding model. You'd probably want Qwen3.8 27B at a minimum, which requires at least 32GB of VRAM (not system RAM) to run semi-comfortably.
So this isn't going to be a competitive model for hobbyists, and you'd have to be a bit desperate to use it for coding. But if you work in a regulated industry and don't mind paying for a bit of extra hardware, it isn't catastrophically bad, either. Probably would work fine for information extraction or as a "classifier" like Jev. (Almost any GGUF model can be turned into a classifier using llama-server. See pi.dev codemode for sample code.)
So they're not a real contender yet, but they look like they're probably at least minimally credible.
Thus training on 'clean' data is like trying to unscramble an egg.
I feel I messed up your quip =/ I'm new here, go ez. Not looking for excuses to hate on China either.
For what it's worth, in my language we don't have a word for copyright either. We have the concept, though, we just call it literally Creators Rights זכויות יוצרים and the borders of what is and what isn't covered broadly map to the familiar concepts of IP.
A real sovereign effort could invest heavily in this, whatever people accuse China of “stealing” I’m sure they are also generating tons of their own data and are probably the primary sovereign doing so outside the US labs.
You're right that especially big models benefit from training on copyrighted material in terms of world knowledge (especially from books). However, in the small model space imho agentic capabilities where the model looks up knowledge on the fly are much more important. That's what we focused on quite a bit during training. Personally, I also don't think stealing stuff is okay.
Regulate large cloud services and proprietary software - yes! But not on the basis of "Intellectual Property".
The "legal" issues here are very very complex and we should not passively wait for or accept corrupt court rulings, international trade agreements, proposed laws, or worst of all propaganda that pushes a parochial and craven view on this.
We as many other’s were curious to try and benchmark it.
On that note, as a small gesture of support, we’ve hosted and made Kolibri-1 free for anyone to try for the next few days.
No GPU. No setup. Just try it. tesseracted.com/kolibri-1-chat/
No, I'm not gonna do that, I asked you to do that Mr Kolibri.
Also feedback on that interface: It's very annoying while answering. It almost immediately shows a list of sources, which on my screen fill up all the space and then when it starts answering it keeps those in view but also scrolls down the tiny part of actual text its outputting but I can't scroll up to start reading from the top, coz it keeps scrolling. I have to wait until it's completely done generating its output.
Uses Brave search, the idea is to test how well the Model can decided when to leverage search or not.
What I couldn't tell but maybe you can tell us: it also complained that it couldn't read the full text as something was cut off. Is that because of the tool you gave it, of brave itself or is it the model?
I wonder how far we are from this. How far are we from LLM's Debian moment?
Also, once the Cohere takeover is complete will they still be able to use this "sovereign" claim despite being 90% owned and 100% operated out of Toronto?
The main problems with big corp AI are due to control of access in the first place and control of what they output.
When you make your industry reliant on such choke points, you render yourself the opposite of "sovereign" for sure. Having multiple independent suppliers at least mediates that.
Given your other comments I can see why it was banned.
I’m most familiar with Canada, where sovereign is usually just an excuse to overpay someone connected for an inferior product with no strategic value.
Also, it's not like it's fixed in time, they are continually building more powerful models as well. Simply because you're in second place doesn't mean you quit the journey. Though, I'm sure the US has a huge vested interest of convincing people to "just quit and submit."
No. Thanks.
Qwen models have solid performance on benchmarks and we're transparent about this in the report. We're just happy to share a European alternative in the small model space, where I don't think we can afford to be fully dependent on China. We also put a lot of effort into German language quality things that don't show up in evals at all (style, Grammar, German reasoning, etc).
Sovereignty is imho mostly about choice and control over your data. Cohere is no different on that front and personally I'm quite excited about what we will build together post-merger.
(It's MoE vs dense, I'm sure. But Qwen's own MoE models don't score as well as their dense models.)
"an LLM" -- does that mean they are effectively learning from that LLM the German encyclopedic style? makes me wonder which LLM and how that is really sovereign.
This thing is worse than a Qwen3.8 27B.
Sovereign works when talking about building a commodity supply or something, not in literally the world’s most competitive and fast moving field.
Those seeking sovereign capability would be better off aiming to be best at something, even something much narrower than an all round LLM. Or just fast following and making something that matches leading performance, which is close to what the Chinese labs do currently.
I think it's not simple to directly compare a 27B dense model with an MoE model like ours. As we know, dense models need all params active for every token. Whereas, MoE models (especially sparse ones like Kolibri) fewer active parameters and correspondingly less compute per token.
Among the MoE models we compared against in our tech report and model card, though, Kolibri performs very well in our evaluation, including against models with 12B active parameters. It also best model in the group we tested within that range of active params.
So, I think it's fairer to see this as a trade-off. Kolibri needs less compute per token but more memory, while Qwen3.8 27B needs far less memory and more compute per token. In the report, both are actually on the quality-vs-serving-cost Pareto frontier among the models we evaluated, just at different points.
https://aleph-alpha.com/en/blog/bounding-hallucinations-merl...
Gemini: Here's a list of links to check
https://minecraft.wiki/w/Minecraft_Support_Virtual_Agent#Sys...
(for context, the Minecraft Support chatbot, aka Merl, had a meme because she kept saying "I don't know" to questions like "How to craft a diamond pickaxe".
That said, I couldn't find any evaluations in the report targeting that specifically. They evaluate on MC for hallucination and look about comparable to Qwen3.5 35B-A3B there.
disclaimer: I‘m part of the training team, happy to answer any questions
- Are there non-LLM approaches to the above task with the goal of achieving a non-hallucinatory agent?
It seems like a very suitable size for local AI models on reasonably high end consumer devices, given it's low active parameter count and a mixed 8bit/4bit quant would fit easily inside 64GB of memory.
Aleph Alpha Kolibri: How the sovereign German LLM works - https://news.ycombinator.com/item?id=49943034 - Oct 2026 (162 comments)
Congrats to the team.
> built for sovereign mission-critical work in regulated areas including public administration, industrials and aerospace
Cool to be doing more independent model lineages, but I sure hope no one actually uses this as part of any aerospace engineering...
More specifically, I asked the model about what I should put on my contact page, which can cost you in the order of 500 € in Germany if you don't write the right magic words.
Kolibri incorrectly referenced the "Telemediengesetz" ("telecommunication act"), which has been superseded by the "Digitale-Dienste-Gesetz" (DDG, "digital services act") since 2024. The model knows about the DDG, but does not reference it unless specifically instructed to do so.
If anyone of the developers reads this, you can fix this by introducing a recency bias during training. You can even control it by conditioning the model on a date provided with the system prompt or first prompt, so you can travel in time.
> trained it on infrastructure in Germany
Made me chuckle. About the second worst place in Europe to train a model if you really care about environmental constraints and energy requirements.
https://app.electricitymaps.com/
Good initiative on the sovereignty front though.
Reiterating what I said yesterday, I think what most people did miss here is that coding is probably not the main goal of this model. It can also code, but the real target is the business processes that aren't code.
Think for example a support department, where new cases (in german!) should be automatically processed (or automatically draft processed) based on existing cases, without pumping all that data through a cloud api.
Or being a general sparring thing for non-IT workers.
__
If you want to code, you already operate in english, so there's the existing and better coding models for you.
But if you're some regular worker somewhere, you might just want to ask the thing to retrieve data on this specific project done three years ago and go through the shared email folder relating to it.
__
You can see that that might be the target, given that they've explicitly focused on getting the model to say "I don't know" instead of just making stuff up. That is exactly the capability you need for these kinds of use-cases.
You can also see that in how it is optimized for speed and a fairly small KV footprint. The idea likely being that a business can pay like a few hundred euros once per worker and then have this kind of local capability for a team of 30.