Kolibri: A Sovereign Open-Weight Model
aleph-alpha.com441 points by bastitx 12 hours ago
441 points by bastitx 12 hours ago
tech report: https://aleph-alpha.com/downloads/tech-report.pdf
additional paper: https://tej.as/blog/aleph-alpha-kolibri
The paper explains absolutely everything as if it was a tutorial "how to made your own modern agentic LLM". They even tell how they made their dataset. https://aleph-alpha.com/downloads/tech-report.pdf ; It's the first time I see this level of openness. I worked on Kolibri, in particular pre-training data and mid-training. We strive to be as open as possible. Glad you like it. How do you cleanse the data at this scale? By various forms of deduplication (exact, fuzzy, substring), heuristic filters and distilling quality classifiers that annotate our data. Synthetic rephrases can also be considered a form of cleaning/getting more out of existing noisy data. We have a lot of details in the tech report if you want to go deeper. Hey! I’m curious if you tried comparing luxical to model2vec classifiers for the pretraining. I’m one of the authors of model2vec, and working on training classifiers for this. I think model2vec could be better, but I haven’t had the opportunity to try this at scale. So if you did, knowing about it would be helpful! Are there plans to make much larger versions of this model? With 500B-1T params for general purpose knowledge tasks, similar to the current leading proprietary models? In my opinion not being open about which data is ingested and trained on, and trying to make that a repeatable thing for a third party, is not worth being called "open". Glad they did that. to be fair, this discussion has been had numerous times here and the industry has arrived on "open-weight" to describe the practice of releasing the post-training weights in an open manner but not releasing the data it was trained on. That's what they call this and I think it's a pretty clear definition these days to people in the industry. I wish universities would take it upon themselves to curate the training sets for these models. The paper mentioned is here: https://tej.as/blog/aleph-alpha-kolibri. (This comment was originally posted to https://news.ycombinator.com/item?id=49943034, but we're merging the threads.) This alone makes it much more valuable than many high-profile releases despite not quite performing at the same level. Such a crazy change from the times of Luminous, when they published a three pager with a claim that the model is similar good as „gpt 3“ (which??) with some graphs without y axis. Bravo team! Hopefully this becomes the new standard. It’s seemed crazy to me that anyone thought these could stay closed or even SHOULD be closed source. Be cautious what you wish for. Tools don't tell you what to do with them. Open source LLMs "democratize" access to the "intelligence booster" that is AI. But while that has several benefits, it also has several downsides. Humanity has the serious problem of being underdeveloped in the "spiritual" department. Ethics is often considered some sort of lifestyle choice, but it's actually the difference between order and chaos in a society. Everybody being able to do anything means somebody will be able to do something you don't like. At an arbitrary scale. Of course openness is only worse than leaving everything under the control of a select cabal of you believe that cabal to be more ethical than the rest of us. The select cabal who believe in an eschaton they're actively trying to bring about, as well. Would you apply this reasoning to the proliferation of nuclear weapons? edit: why is the parent rationale sensible for AI and not nuclear technology? Absurd and incomparable. Why? That seems like a shallow dismissal. Why is a world changing technology okay in the hands of a small cabal in the one case and not another? "Why is the wheel not okay to gatekeep but the nuclear bomb is?" Even the framing is manipulative from the beginning. By asking this question you're already assuming they're in any way comparable. They are not even remotely equivalent and the entire comparison is utterly absurd. AI is more like a nuke than a wheel. Do you disagree? Can you suggest a less manipulative framing? I am personally a proponent of open source AI and models but I found this cabal framing strange when we do indeed rely on this kind of control for other world-altering technologies. And AI is different in that it enables technological development in ways quite unlike the wheel in a general sense. The main thing that bugs me about the risks discussion around AI is the lack of specificity. Commenters here have a good grasp of the risks around finding vulnerabilities faster than they can be patched. That's good and it matches the applicability of LLMs to coding. But the applicability and the ROI of LLMs for other use cases than coding is a lot squishier. Also correspondingly the risks are unspecific. As for what to do, ethical disclosure of vulnerabilities provided a good framework for disclosing software vulnerabilities discovered with the assistance of LLMs. What is going to be novel and calls for our spiritual development in other domains? I find it odd that more people don't realize that we are talking about risk of elevated *general intellectual-domain capabilities* as a resource. To be clear, I'm not claiming that LLM + RF is necessarily THE technology that poses the risk, the risk is in recursive self-improvement and whatever technologies will result. It's very clear to that any specifics couldn't capture the risks, because the capabilities, including the risks, are one level higher than any specific techonolgy. It is the process of advancing technology itself, in accelerating speed, that poses the risk. The reason I use coding and vulnerabilities as an example is that it is a concrete example. It is what people pay for now when they buy AI. And the risks are specific and can be examined in detail. Some threads on this board currently show that even these more concrete and specific risks are often overblown, with LLMs finding low risk bugs and sucking up resources to evaluate and fix them. Here you are claiming that AI products are going to reach AGI or RSI in the foreseeable future. Of course you can't "capture the risks" with specifics because those are inherently unspecific futures. It's a bit like saying when we invent antigravity all hell will break loose. I would believe those future risks more if there were a progression of risks. What other than finding vulns has those characteristics? What poses the risk is the combination of abilities past a certain point enabling you to do basically anything. While being unable to judge whether you should in the first place. ai cant move atoms at unlimited rate and also have limited energy. so your claim that "basically anything" is a bit of a stretch. You're right, people weirdly lack imagination on what "higher intelligence" (minus ethics) actually affords you, let's have a look: What do average people currently want? They're taught, the most important thing was being rich. So they will ask their AI to make them rich. Most real life ways to get there are "sketchy" to say the least, usually downright unethical and anti-social, but US society turns a blind eye when the "Wolf of Wall Street" comes out on top and the schemes don't easily fit into average people's abilities of moral judgement. -> Large parts of US society suddenly engaging in all kinds of "semi-legal/hyper-illegal" fraud schemes, at the expense of already saturated environmental and societal resilience. Guaranteed collapse. Or, let's get rid of those pesky neighbors/wrong-colored people/annoying opinions? Again, "legal" is a pretty squishy concept and only really applies when you don't have the legal expertise to get around it. Now you can. Or, look at the basics: what is "real"? You only "know" because you trust certain people and institutions. Generative AI can help with that /s. It's not only about "building weapons of mass destruction". It's about doing the same shit as usual, but a thousand times faster/amplified. Look up poly-/metacrisis for starters. Going faster with AI when there's a wall in front of you isn't the best idea. This is analogous to the problem of spam, which is a problem about five minutes younger than email. Before spam you had to buy ads in the back pages of magazines you think target vulnerable demographics. Meta already spews fraudulent ads in horrific volume. In other words, ambitious frauds have already explored all of the angles and bought all the ads. At worst, LLMs will create a few more successful but less ingenious frauds. Whenever classic p(doom) sentiments are explained through a text with obvious LLM markers, I wonder whether I am looking at a superhuman persuasion attempt. ..or maybe the commenter did look at superhuman persuasion long enough to believe it would be best to channel those ever the same fear fantasies from the LLM through their account to the reader. On a more serious note, just look at the doom premises here: "Large parts of US society suddenly going criminal" is from the movie "The Purge", I think. It is fiction. The idea that generative AI takes away our ability to find out reality. ... I don't know. People write about that a lot, but it still seems very far fetched. Maybe through some terminally online overconsumption, like with social media? I wouldn't know. With new AI capabilities we will have to adjust, I am sure. Media, science, education and law are changing very visibly right now. Those p(doom) narrations just seem to be pre-IPO hype though. It is just so so dangerous. That is why they want to go public and only want to care about optimizing for the next quarter ...right before breaking into AGI. /s Ah, didn't think of this. Some rogue militia group might try to use this LLM (or create their own LLM based on this work) to help them create biological weapons or to do a mass hacking the infrastructure of targeted country. Wonder what safeguards Kolibri uses to prevent this? Or if they even can I don't think that AI is as much of a boost to bioweapons as people think. Lab work doesn't get easier just because the experiment design part does. Chemical and biological warfare is hard. A cult in Japan created a mass casualty event using nerve gas. Which is the only somewhat "successful" terrorist WMD attack I know of. Knowledge of how to create these weapons isn't new. Guns and bombs are the most widely used terror weapons for a reason. Enter LLMs to help through all those pesky hard parts. The hard part is probably finding the lab equipment and chemicals without being noticed, and choosing not to use it to manufacture drugs instead which would be much more profitable Intelligence was never the bottleneck tho For terrorists it is tho. Four lions is basically a documentary. The only clever and creative terror attack happened 25 years ago. Didn't Mexican cartels kidnap telco workers to build them separate infra? We expect terrorists to be less resourceful? What's stopping them from using claude today? What could anyone possibly do to stop them from using open models? This line of thought seems like corporate/political/pr pandering more than a meaningful concern. Oh definitely, and I’m in the cybersecurity space so I’m already on the “worst case scenario committee” hah. But the alternative just seems… so much worse to me? A select few groups gating access to the ability to do everything seems like neo-fuedalism in the making. And to be fair even the gating that we do have (daybreak, CVP, etc.) is already being circumvented via keys being stolen and sold on the dark web. Yes, a "select" (rather, self-selected) elite "controlling" AI according to their wishes, what could go wrong? Clearly not a "better" scenario. The real problem though seems people feigning helplessness? You can't leave society "to its own". You are part of it and go where it goes. So better start steering. When access to AI gives you abilities you cannot use responsibly, you shouldn't have access to that. Just like you shouldn't be allowed to drive a car or fly a plane or command a rocket without proper guardrails, safeguards, prerequisites, etc. "General" intelligence isn't present in humans, why does it need to be in AI? > When access to AI gives you abilities you cannot use responsibly, you shouldn't have access to that. What do you mean "cannot"? As in you are granted abilities that have no responsible use? > When access to AI gives you abilities you cannot use responsibly I don't think this is realistically a problem at all. It just makes certain types of research cheaper and less time-consuming. And, again, this is also a problem with the american services. I think the unethical things are happening already with boutique firms. I admit I don't get the concern. > Humanity has the serious problem of being underdeveloped in the "spiritual" department. As is shown to us by filthy rich people every day. Or did you mean the burglar in the fawellas? His point about regulation and innovation is great and I wish more people thought like that. One of humanity’s biggest problems here is we don’t know how to do moderation. We have two modes. One is a brick taped to the accelerator and damn all consequences, driven by national pride or corporate greed or egos. The other is a brick taped to the brake driven by histrionic doomers and anti-everything pessimists. The extremes are loud and fit in a tweet. Nuance is quiet and contemplative and usually requires an essay or a book. It’s also dynamic. Nuanced positions evolve over time as new things are learned. Extremes tend to be fixed and rigid. All this, I think, gives them higher memetic fitness in the discourse. I don’t think this is new. Look at nuclear power, a largely pre-Internet example. You had pro nukes who minimized and hand waved away any risk and anti nukes that wanted it utterly outlawed. Nobody said “hey this is a great zero carbon source of energy but we really need to think it through carefully and manage it well.” Or if they did they were drowned out by the loud screaming extremes. (This comment was originally posted to https://news.ycombinator.com/item?id=49943034, but we've since merged the threads, so I've moved it into the subthread which is specifically about the paper being responded to.) >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. https://aleph-alpha.com/en/blog/bounding-hallucinations-merl... I'm not super impressed, at least in English. I asked: Just like Merl. 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". Gemini gives me a lot of hard-no responses, and I think to myself, "well can you find out? what am i paying you for" usually Gemini: Here's a list of links to check Thank you Aleph Alpha team for making it open. 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/ Friendly note: your website's font at its current size is pretty bad on a non-retina display before zooming in. It's been surprisingly good at the things I threw at it (history, culture and conversing, web search)! 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. Correct, search is added as a tool. Uses Brave search, the idea is to test how well the Model can decided when to leverage search or not. 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? Thank you for the feedback on UI, improved the streaming to make it less frustrating. 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? For a post to make such a big deal about sovereignty it is a bit misleading to not mention that the company is slated to be merged with Cohere, a Canadian company. And that is a good thing - no need to hide it. Given the growing cost of keeping up, these few non-US, non-Chinese companies really need to do more sharing of efforts and costs. Canada too is very much in need of sovereign AI options, but funding that on its own would be pretty much a waste of money. Would love to see this new German Canadian company cooperate with Mistral too, or maybe one of the Korean AI companies. Given that "sovereign" these days implicitly means independence from USA and China, and Canada being spiritually in the same boat as the EU, it's pretty spot on? The costs of "keeping up" aren't really growing, on the contrary. I also read about Canada maybe joining the EU in some capacity, although I'm not sure if that was a joke. Eurovision could be a start. Ursula von der Leyen (president of european commission) propsed an "associated membership" with EU. Such an association is not defined anywhere in EU contracts yet. https://www.iss.europa.eu/publications/commentary/eus-first-... Hey there, I worked on Kolibri pre-training data. Agree with the point that non-US/non-Chinese labs need to pool effort. The merger with Cohere is public, nothing to hide and I'm also excited about it personally. 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? 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. 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. 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. My account is banned, so for whomever with [showdead] on: What's so "not-sovereign" for an open-weights American or an open-weight-open-training-process Chinese LLM? Do we also need sovereign Linux (maybe), sovereign Postgres (most likely not), sovereign Python (def not)? If it's banned then maybe you shouldn't be posting... Given your other comments I can see why it was banned. I was also quite surprised to see that. Considering the effort Aleph Alpha put into to their new model, it seems like the Qwen team needs access to vast amounts of german data o.0
I'm really glad for these efforts for open models from within Europe. I think at the moment the main thing a sovereign AI model needs to be good at is auditing the results of other models. Right now one could run an open model for most government applications and it would be good enough, you just cannot trust any of these. So having a sovereign controlled model audit the first one would basically act like a “trust adapter”. If the second model is cheap and fast enough, there is a business model. You don’t even need to audit all the intermediate steps, just tool calls and end results. The issue with 'sovereign' is not about sneaky things a foreign entity might put in there, it's about control of the stack. Since no individual buyer cares that much about geostrategic issues, it's almost impossible to get some random European company to think about buying anything other than 'whatever US or China' are making. There has to be very concerted push to make a difference. Legal mandates around sovereignty (be very careful here) could make a difference. But there will also be market reaction: US/China companies will bend on some level and provide things like 100% EU hosted and even comply with some data source stuff. The only real path is to be more competitive as a continent. If you self-host an open model on a sovereign hoster? They can only influence the responses correct? We are not talking about extracting information. Or are we worried that open models send secret telemetry? We're talking about owning the stack - who gets to use what text, for what reasons, under what circumstances etc. Competitors, other nations, Russia / China, the 'global south', military contractors etc. etc.. The EU AI act isn't primarily about "geostrategy". It's about protecting society. AI models act as a force multiplier for intelligence, in particular for generating information according to someone's wishes. I.e. deepfakes and social media mis-/disinformation campaigns are a thing and having powerful AI allows you to do those at scales that can overwhelm society's resilience. In general, even if you have "aligned" AI: aligned with whom or what? Whom are you comfortable with lording as a some demi-god over you, dictating what to believe? European regulations are 100% geostrategy and they are defending against external interests. Those laws would not exist if European champions were leading the world. "are you comfortable with lording as a some demi-god over you, dictating what to believe?" Yes, Europe handed over all of the decisions about everything to foreign powers, now they have to enact regulations to try to constrain it. Zuck et. al. make the investments decisions for Europe, by virtue of you all giving him the money and power to do that. Stop giving him the money/power, then this regulation won't exist > Since no individual buyer cares that much about geostrategic issues You clearly haven’t worked in enterprise in the EU. Location of data processor is the first thing they check. That’s why every major cloud has regions with different offerings, not just for HA and redundancy They care because the are 'required to', otherwise those 'offerings' wouldn't even exist. It's mostly the result of regulatory action. I’m working with clients (big companies) that care a lot - but not because they have to. They care because they lost trust in non EU players. They also care because it's becoming increasingly risky to rely on US companies given the current political situation nope, that no longer holds any water. was true though couple of years ago… The argument holds most of the water. Yes - security concerns are now very real, but do not fall for this idea that anyone really cares about structural concerns. Tons of EU companies are still selling crap to Russia, happy to look the other way wile the bear devours a neighbour, as long as profits are there. What has happened is that Trump has given a face to the reality, and so companies are adjusting on some level. But 1) it will be nominal 2) Trump will be gone and the impetus will fade 3) the conglomerates will react 4) modest legislation will mean ... The can will get kicked down the road. The invasion of E. Europe by Russia has not even caused defence spending or the size of Armies to change, doctrine is barely changing. Nothing will make those beheamoths change other than other forms of structural concern. The 'Riet of the Right' - partly due to collapse of Auto / China (among many other things) might cause some change. But the change won't happen until well after the damage has been done. Brexit should have been an opportunity for institutional reform of the EU, but no - they blamed it all on populism. Certain governing entities in Europe will try to move away from Microsoft, they will be pulled back. There is hope, and certain champions can rise and causes pieces to collapse. If SUSE had any true entrepreneurial whereiwthal - they would create a true consumer / prosumer / enterprise-user friendly variation and brand their flavour of Unix - and make sure that all Euopean governments us it exclusively, which would cascade into widespread use. There are variations of insta, youtube, netflix that should all be Europe based that could 'theoretically happen' but it needs some structural impetus. Things usually don't change. Usually there needs to be a collapse and re-order of the system for that to happne. The people in charge just want to keep their jobs, their very high salaries, and protect their retirement, and will be happy to 'sell out' whatever other imperatives along the way. Ending with a positive not - I would say current conditions mean the change is now 'plausible, it not likely' whereas before it was 'not very plausible'. So there's a candle, a bit of wind, but not a lot of tinder or dry wood. That is a very good idea and there is a market for that, especially if a company manages to source the hardware and then install the stack on a german or EU customer site It depends on the use case, but I believe SLMs can do great with smaller tasks. People got so attached to 'general purpose' that they forgot software can be designed for specific, smaller use cases. The absence of any comparison to Qwen3.8 Flash, another MoE model with a small-ish (6B) number of active parameters, is pretty striking. Instead, it's compared with Qwen3-Next 80B-A3B, a model released almost a full year ago. I get that doesn't invalidate the real "point" of the model, but... That is what stood out to me as well. Qwen 3.8 Flash can run on very limited hardware as well and it is at least as good as Sonnet 5 in benchmarks like DeepSWE. With its n-gram design you can get flash next running on very limited GPU resources, as little as 16GB of VRAM. People follow the latest frontier lab models with great attention and migrate to the next big model on their subscriptions. Meanwhile these local models have quietly gotten REALLY good. It is not even an exaggeration. It has happened in the last couple of months. "Local model you can run at 40 t/s on a gaming machine that is better than Opus 4.6" is way less exciting than "OpenAI IS DOING CRIME!!! OpenAI SOLVED NAVIER STOKES. DARIO SAYS GLM 5.3 BAD! SLOW DOWN THE FRONTIER!". (edit: also... totally ignore that 27B dense column over there where Qwen 3.8 27B beats Kolibri on nearly every single benchmark. Why would I choose to run this model?) I think the simple reality is that if your goal is producing quality work output, you want the smartest possible model available. What is the upper bound on the value of more intelligence applied to your problem domain? The point of a model like this is aimed at providing local inference. You want the smartest model available, as long as it meets all of your other criteria. That isn't always maximizing on the absolute smartest model. There are many reasons to avoid a frontier model for now for a variety of reasons. All of this is even only relevant in the last 6 months anyhow. So it isn't like this is even some long term trade off or position I am proposing. Can't you infer the comparisons you would like from baseline results provided? There are better results in the sibling post on HN: https://aleph-alpha.com/en/blog/kolibri-has-landed-a-soverei... Maybe? Qwen3.8 27B scored notably higher in most of the provided benchmarks, including the German-specific ones. The only "downside" is that inference is much more costly and slow, since it's a dense model. Qwen3.8 Flash-Next appears to usually "benchmark higher" than 27B, while remaining fast. I'm sure I could dig up the equivalent benchmarks for Flash and do the comparison myself, but as far as inference goes, it's messy. Consider that Qwen3.5 35B-A3B scores higher than Qwen3.6 on some of the German-specific benchmarks. So it seems superficially plausible that Qwen3.8 Flash-Next might not be "27B but faster" in the ways that are important for this model. Or it could just "be superior" in all ways. Either way, I don't think an LLM has to be "the best" at anything to be worthwhile, necessarily. And I kind of distrust benchmarks on top of that, so... I just tried to play around with it on my RTX pro 6000 setup, it spends way too many tokens on overthinking stuff even if it’s able to catch the correct approach Its speed is pretty good on the other hand with only 3B active parameters I am getting around 170 tkn/s on fp8 > it spends way too many tokens on overthinking stuff Yeah. It’s a German model. This joke is either gonna get over analyzed by the Germans or gonna fly past straight over their heads I'm not sure if this is truly a good faith joke tbh. [flagged] No, I am just standing my (and my nations) ground against what is simply bad faith disrespect and what could be described as casual racism (or at least spiritually equivalent to that, to not get into the racism definition debate). You do not get to invalidate that. People need to stop insulting my nation. __ If the people you're joking with aren't laughing, you're not joking with them but about them, which is generally seen as a bad move, and something that we actually wanted to leave behind with all the social progress and all. __ This really only happens because people (as all bullies) think that they would be free to do so, because the victims would not fight back. (Un)fortunately though, the world now has bigger problems with genuine fascists, so neurotically cowering in fear and performatively saying "yes, punch me harder. Insult me more" can be stopped now. __ I mean look at it. By now, it's at least better, but the first 30 comments or so were "haha germany bad haaa". That's a terrible showing for the platform even if you're not german. What value was offered? Who would want this. We don't even need to talk about nations and identities to see that that was just noise. Nothing flies anywhere, we use superior high speed trains. I faxed a Humorgenehmigungsantragantrag to our Bundesunterhaltungsministerium analyst, immediately. Laugh now, but you are merely lucky the reply got delayed. As soon as the leaves on the rails are dealt with, you will get a formal response that has washed itself! Let us see who is laughing then. It is nobody! But, most importantly, is Bundesunterhaltungsministerium a single token in Kolibri? This could make a huge difference for the model's performance in German. Silly you, thinking the German economy is merit-based. The only question is, if the bribe fits a single token of appreciation, to win a tailored call for bids! petition to make this thread the official german jokes thread for this post, it's kind of annoying how everyone feels the need to make a top level comment for their oh-so-great "germans suck" joke. Related ongoing thread: Aleph Alpha Kolibri: How the sovereign German LLM works - https://news.ycombinator.com/item?id=49943034 - Oct 2026 (162 comments) I love how the mere mention of a "sovereign" in LLM's announcement is the declaration of defeat. This thing is worse than a Qwen3.8 27B. 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. 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. 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. Looks interesting. I just went to download from HF, but they only have fp16 which won't fit on my Mac. Good to see Europe adding toe what Mistral is doing. +100 FP8 weights are actually the default and what we trained for: http://huggingface.co/Aleph-Alpha/Kolibri-1 We have a separate repo for BF16: Kolibri-1-BF16. Will still be tough to put it on a Mac though :( Disclaimer: I'm part of the team that trained Kolirbi "Languages: German and English" – this is odd. That means their dataset is limited. In my understanding, frontier models are trained on multilingual datasets and can combine knowledge no matter what language it was written in. Have you read the non-English output of those mutlilingual models? They say they put extra effort in to improve nuances and tone in this one. That alone is worth having in many settings. Related thread on the tooling: Model Training as Code - https://news.ycombinator.com/item?id=48673450 - June 2026 (24 comments) And related thread on the actual model post: https://news.ycombinator.com/item?id=49942706 > 3B active > 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... 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 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. For those sick of “Pareto frontier” talk, just shorthand it as “it’s the best at some very particular thing”. Obviously that one thing/tradeoff it’s good at may not necessarily be compelling, but it is either a loose sign of quality, or a sign that they’ve chased some tiny edge into the ground. I’ll be curious to see which it becomes in the next year - nba “very narrow record”, or a sign you can hang with the big boys. Was expecting that a "sovereign" AI model would at least use their own sovereign language (German) on the website as one of the options. Anyways, all the best and happy reunification day. This is a blog post by some "random" dude right? The company webpage appears to be German (at least when accessed from Germany: https://aleph-alpha.com/) The blog is from the company itself, just noticing the irony. > 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. I wish nothing but luck for an EU model, but: > 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. 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. Edit: reading through the full text of their post, it looks like they are using common crawl, which is likely just as much of a copyright infringement as Anna's Archive -- it's not like published works have a unique claim to copyright. I think this strengthens your point, though: I was expecting to see scans as training data, but it doesn't appear to be the case. "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. 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. I hope you're right, but I guess we'll see when we get independent benchmarks. My intuition is that even for small models and models mostly focused on tool calling, you _still_ need all that extra contextual stuff for the magic, but I am further from the coalface than you are. 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. 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. There is no word for copyright in Mandarin :) 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. 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. 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. Have you tried the model itself and seen if it's "even slightly competitive" or not, and have specific complaints about it? Otherwise it feels like you're complaining about something that is easy to test but rather than taking the time to actually figuring that out first, you're arguing about some general and theoretical thing which the submission (may) directly disprove. 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. 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. That's fair, but then people like you complain when someone "steals" I mean distills openai or anthropic models. It’s poor form to argue against someone by imagining something totally different that they might believe, which would then make them hypocritical. 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." Congrats to the team. 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 wonder if the custom tokenizer is better in practice, the examples look interesting though 3.5b active params sounds cheap until you remember all 78b still has to fit in memory. curious what the smallest practical self-hosted setup looks like for german docs. The benchmarks are impressive given the problem space they are working in. German here. We are cheering for Mistral, which is making some good moves before the year is over, and Black Forest Labs for non-coding. But that's about it. Im surprised by how well this works. What is the difference from this and union alpha (other than the fact that it is open weights)? is this like an unfortunate name clash with KolibriOS (kind of like how Google Gemini was a clash with the Gemini protocol project)? 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? Is this a truely open source model or also open weights? The weights have been published to HF on an apache2.0 license, and the tech report is quite extensive > A bigger dense model beats it. Qwen3.8 27B ... How is a 27B dense model bigger than a 78B MoE? The 27B has all 27B as active parameters, while Kolibri is only 3.46B active parameters. I think that's what they meant anyway. I got distracted by that scroll-wheel UI component on the page. Neat! Can somebody mention the URL of the page with this? We're merging the threads and I don't want a dangling pointer! The scroll wheel was cool indeed. Check their home page.. their talks on a carousel over a horizontal timeline was cool even on a phone browser. 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. It still steals my IP without attribution. Now we have state sanctioned sovereign theft instead of foreign theft. Aleph Alpha is just a sad joke by now. The talent isn't there anymore, they never managed to catch up to the other labs, failed to deliver on several projects, and by now are just a cash grab for the investors. Only if you think in terms of short term gains, but thinking in the long run, doesn't really matter if these models are crap today, all models will eventually be obsolete, unless we plateau hard on every aspect of the tech. What matters is to have good sovereign models in 5, 10, 20 years time. I'm curious why you think sovereign models matter that much if open source ones exist (unless that stops at some point). Sovereign model hosting services, yes. Just open weights are not enough, I think as a nation state you'd at least want a whole sovereign training pipeline. Otherwise you are stuck when Kimi K4 stops releasing their weights. But I don't see why this needs to be done on a state-by-state level, a sovereign EU model seems to be much more realistic in terms of funding. Wouldn't you also want to be able to manufacture some sort of GPU? I actually think we should be revisiting PMT photomultipler tube construction, and electron-"optical" matrix multiplication. The annoying part of optical matrix multiplication is the conversion from electronic to optic and back. Why not use electron-optics for matrix multiplication and use dynodes for amplification. While an intriguing concept, that doesn't sound like it could scale to the amount of "cores" needed to do the work. Or am I missing something? We have lots of capable open weights models, but very few have actual datasets and pipelines open for reproducibility. This is a major issue that even if big companies addressed, policies could well just change, as you also acknowledge. Society can't depend on the whims of the private sector. > Society can't depend on the whims of the private sector. Yeah we can, we do it all the time. Nothing in society works without relying on private actors to keep doing what they have always done. Unless you go full soviet command economy, which in fact works worse. Society and the private sector cooperates, but any technical dependency is detrimental because of monopolization of interests. You don't need to go full soviet, but knowledge in general must be socialized, that includes access to technology. We literally grant 20 year monopolies on technologies to their inventors and have for centuries and it's fine. The entire economy runs on chips made by like, 3 foundries using lithography machines made by 1 company. I think you are overstating the necessity. This example proves the exact opposite, that is a national security disaster waiting to happen Autarky would be a national security disaster manifested. You can't be a modern prosperous society and not rely on either trade nor private enterprise.
miellaby - 7 hours ago
ivo-42 - 4 hours ago
brcmthrowaway - 4 hours ago
ivo-42 - 4 hours ago
stephantul - 3 hours ago
xvfLJfx9 - 2 hours ago
zelphirkalt - 7 hours ago
davidjfelix - 43 minutes ago
amelius - an hour ago
dang - 4 hours ago
p-e-w - 6 hours ago
zwaps - 4 hours ago
ofjcihen - 6 hours ago
Loquebantur - 6 hours ago
adrianN - 5 hours ago
zanderwohl - 4 hours ago
skinfaxi - an hour ago
orbital-decay - an hour ago
skinfaxi - 42 minutes ago
orbital-decay - 34 minutes ago
skinfaxi - 23 minutes ago
Zigurd - 5 hours ago
GolDDranks - 4 hours ago
Zigurd - 3 hours ago
Loquebantur - 4 hours ago
vincnetas - 3 hours ago
Loquebantur - 5 hours ago
Zigurd - 3 hours ago
fwn - 3 hours ago
computerdork - 5 hours ago
zanderwohl - 4 hours ago
Zigurd - 5 hours ago
UberFly - 5 hours ago
shawabawa3 - 4 hours ago
throwaway27448 - 3 hours ago
happosai - 2 hours ago
skinfaxi - an hour ago
throwaway27448 - an hour ago
ofjcihen - 5 hours ago
Loquebantur - 4 hours ago
skinfaxi - an hour ago
throwaway27448 - 3 hours ago
throwaway27448 - 3 hours ago
OtomotO - 4 hours ago
api - 9 hours ago
dang - 4 hours ago
andai - 4 hours ago
Glitch752 - an hour ago
(A made-up song and name)
And received: What is the canonical interpretation of the song ‘Glass Flowers’ by Armand Uso?
I saw similar responses for other questions. Qwen3.8-27B correctly refused without web search. "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. [...]
Velocifyer - 3 hours ago
swozey - 4 hours ago
kkm - 8 hours ago
trvz - 6 hours ago
moritzwarhier - 4 hours ago
tharkun__ - 6 hours ago
sigmar - 4 hours ago
kkm - 3 hours ago
tharkun__ - 2 hours ago
kkm - 5 hours ago
peterBlue75 - 4 hours ago
ducktective - 3 hours ago
tomComb - 6 hours ago
Loquebantur - 6 hours ago
user_of_the_wek - 3 hours ago
mkesper - 3 hours ago
ivo-42 - 5 hours ago
amoshebb - 7 hours ago
befelix - 4 hours ago
slow_typist - 40 minutes ago
andy99 - 6 hours ago
Loquebantur - 6 hours ago
bewareofscams - 5 hours ago
satvikpendem - 2 hours ago
isusmelj - 7 hours ago
niemandhier - 8 hours ago
holaysuns - 8 hours ago
jwpapi - 3 hours ago
holaysuns - 29 minutes ago
Loquebantur - 6 hours ago
holaysuns - 5 hours ago
Oras - 8 hours ago
holaysuns - 8 hours ago
JaggerJo - 7 hours ago
rokkamokka - 7 hours ago
bdangubic - 7 hours ago
holaysuns - 5 hours ago
mawadev - 7 hours ago
adishiktensward - 5 hours ago
spijdar - 8 hours ago
bitexploder - 5 hours ago
frumplestlatz - 3 hours ago
bitexploder - 3 hours ago
okamiueru - 5 hours ago
spijdar - 5 hours ago
martianvoid - 9 hours ago
gizajob - 9 hours ago
sajithdilshan - 9 hours ago
hypfer - 9 hours ago
jamiek88 - 4 hours ago
hypfer - 4 hours ago
jijijijij - 6 hours ago
fph - 3 hours ago
jijijijij - an hour ago
dudefeliciano - 8 hours ago
dang - 4 hours ago
tornikeo - 6 hours ago
tfburns - 2 hours ago
andy99 - 5 hours ago
okamiueru - 5 hours ago
gpugreg - an hour ago
mark_l_watson - 6 hours ago
befelix - 3 hours ago
vzaliva - 7 hours ago
AnonymousPlanet - 2 hours ago
peterBlue75 - 2 hours ago
satvikpendem - 2 hours ago
thatguysaguy - 2 hours ago
lmf4lol - 36 minutes ago
james45 - 7 hours ago
kouunji - 4 hours ago
cheesecakegood - 6 hours ago
x1watt - 8 hours ago
flohofwoe - 8 hours ago
x1watt - 2 hours ago
dosinga - 7 hours ago
petesergeant - 8 hours ago
rpdillon - 8 hours ago
mapontosevenths - 6 hours ago
befelix - 4 hours ago
petesergeant - 4 hours ago
jamienk - 3 hours ago
ekidd - 8 hours ago
torginus - 7 hours ago
miohtama - 7 hours ago
dotancohen - 5 hours ago
cheez-wiz - 5 hours ago
andy99 - 6 hours ago
embedding-shape - 8 hours ago
petesergeant - 8 hours ago
ygjb - 7 hours ago
Zambyte - 8 hours ago
folkrav - 7 hours ago
pepperoni_pizza - 8 hours ago
brookst - 7 hours ago
wg0 - 3 hours ago
peterBlue75 - 2 hours ago
tosh - 8 hours ago
Jeeetendra - 6 hours ago
pythonic_hell - 9 hours ago
d2kx - 9 hours ago
CorezIoOfficial - 8 hours ago
erelong - 5 hours ago
oblio - 7 hours ago
JaggerJo - 6 hours ago
peterBlue75 - 4 hours ago
woadwarrior01 - 8 hours ago
SyneRyder - 8 hours ago
cbarrick - 9 hours ago
dang - 4 hours ago
reacharavindh - 8 hours ago
cyanydeez - 9 hours ago
12949468 - 6 hours ago
9dev - 9 hours ago
gchamonlive - 9 hours ago
roncesvalles - 8 hours ago
HotHotLava - 8 hours ago
rixed - 7 hours ago
DoctorOetker - 6 hours ago
pstuart - 5 hours ago
gchamonlive - 8 hours ago
brainwad - 6 hours ago
gchamonlive - 6 hours ago
brainwad - 5 hours ago
gchamonlive - 5 hours ago
brainwad - 4 hours ago