A 78-Billion-Parameter Model You Can Self-Host: Kolibri Lands With Apache 2.0 and a German Accent

On October 3, 2026, the German AI company Aleph Alpha published the full weights of a new language model called Kolibri on Hugging Face, free for anyone to download under the Apache 2.0 license. It is a 78.1-billion-parameter model that understands English and German, handles up to a million tokens of context, and, crucially, runs entirely on hardware you control. For the self-hosting community, it is the most interesting model release of the year so far. Not because it is the smartest model anyone has shipped, but because it is the most capable one you can run without sending a single prompt to someone else's servers.

A big model with a small engine

Kolibri is a mixture-of-experts model, or MoE. Most chatbots you have used are dense: every parameter participates in writing every word. Kolibri splits its 78.1 billion parameters into 384 specialist networks, called experts, inside each of its 50 layers. For every token the model reads or writes, a small router picks the six experts best suited to that token, plus one shared expert that always runs. The other 378 sit idle for that token. Each token therefore costs roughly what a 3.46-billion-parameter model would cost in compute. That is the hummingbird idea behind the name: a large model that moves only a small part of itself at a time.

The practical consequence is speed. Doing the work of a 3.46B model per token means Kolibri generates text quickly on hardware that would crawl with a dense 70B model. Aleph Alpha reports competitive benchmark scores: 96.9 on the AIME 2025 math competition in English (87.5 in German), 84.3 on GPQA Diamond, a hard science test, and 85.9 on LiveCodeBench, which tests writing code. Those are the company's own numbers, not independently reproduced. On agentic work the picture flips: on SWE-Bench Verified, which tests fixing real bugs in real repositories, Kolibri scores 66.4, trailing rivals built for tool calling. It is a reasoner, not an autonomous coding agent.

MoE saves compute, not memory. All 78 billion parameters must stay loaded, because the router can pick any expert for the next token.

The memory bill is the real price

Here is the catch that will decide whether Kolibri ever runs in your homelab. The router can choose any expert for the next token, which means every expert must be loaded at all times. You pay the compute of a 3.46B model, but you pay the memory of a 78B one. The official FP8 weights take about 78 GB of GPU memory, plus room for the conversation cache. Aleph Alpha's minimum hardware list reads like a datacenter shopping list: two A100 80 GB cards, two H100s, or a single H200, B200, or B300. The recommendation is even more generous.

That rules out nearly every homelab, but the community has not waited for permission. Within days of release, community builds brought Kolibri down to consumer hardware: MLX builds run it on a Mac with 36 GB or more of memory, and a patched version of llama.cpp runs it on an ordinary PC, even on the CPU alone with 64 GB of RAM, at a usable 12 to 15 words per second on a desktop processor with 128 GB of memory. One tester reported more than 50 words per second on a 2021 M1 Max MacBook Pro. The catch is not speed, as one guide put it. The catch is memory: you still have to fit all 78 billion parameters in, even though only a sliver works at a time. A community Ollama upload exists but is unofficial, and stock Ollama, LM Studio, and llama.cpp did not support the model on release day.

What sovereign actually means here

Aleph Alpha calls Kolibri a sovereign model, and the word deserves a plain reading. It means you download the weights, run them on hardware you choose, and no request leaves your network. Nobody can change the model under you, raise the price, or switch it off. The model was trained in Germany and Finland under European and German law, Aleph Alpha has signed the EU's General-Purpose AI Code of Practice, and the company says it built Kolibri with the EU AI Act and GDPR in mind. For a European clinic, municipality, or engineering firm whose data cannot legally leave its own infrastructure, that is a concrete compliance story, not marketing.

Sovereignty here is about control, not purity. The model card acknowledges that some synthetic training data was rephrased using non-European models, so the claim is about who operates the model, not where every input came from. It is to Aleph Alpha's credit that they wrote that down. And there is corporate context: in April 2026, the Canadian company Cohere announced plans to combine with Aleph Alpha, with a definitive agreement signed on September 16. The deal still needs regulatory approval. An open-weight release with a permissive license is one way to make a model's future less dependent on any single company's fate.

The license, in plain language

Kolibri is open-weight under Apache 2.0, not fully open source. The weights and configuration files are yours to download, run, modify, and ship in commercial products. That alone is a real difference from models released under community or research licenses with user caps or non-commercial clauses. The fine print: Apache 2.0 covers the weights and config files only. Aleph Alpha keeps the rights to its training code, model architecture, parameter settings, and training methods. There is also a published list of prohibited uses, including illegal activity, weapons, and malicious code. Whether that list binds you is a question for your lawyer, but if your use is on it, Kolibri is the wrong tool anyway.

Should you run it?

Run it if you have, or can rent, an H200 or B200 class machine, your users work in English or German, and your data cannot leave your own infrastructure. It fits best for document search, internal assistants, and German legal or compliance work, where it was also trained to abstain from answering when the evidence is not in the provided context, a feature Aleph Alpha calls the Merlin-Arthur protocol. Skip it if you are on typical homelab hardware, need more languages (only English and German are tuned), want image input (it is text only), or mainly want an autonomous coding agent, where the tool-calling numbers say a rival like Qwen3.6 serves you better. Either way, Kolibri sets a precedent worth watching: a serious, regulation-aware model released with weights you can actually own. The hummingbird is small on purpose. The idea behind it is not.