Aleph Alpha has released Kolibri, an open-weight language model aimed squarely at organizations that want to run AI on their own hardware. It is an English-German mixture of experts: 78 billion total parameters, 3 billion active per token, context windows up to a million tokens, and full weights on Hugging Face under Apache 2.0. The Heidelberg company’s framing is sovereignty: public administration, industrial firms, aerospace. In practice that means on-premise deployment, no data leaving the building, and a supply chain someone can audit.
What it is, concretely
Kolibri is served through vLLM with a plugin the company publishes, either as a container image or a pip package that pins its own vLLM version. The release notes give a working serve command with reasoning and tool-calling parsers enabled, and note that contexts beyond 262,144 tokens need an explicit override flag. They also publish the sampling parameters they recommend, which is the kind of detail that saves an afternoon.
The benchmarks are self-reported, and they cover the axes this audience cares about: AIME 2025 at 96.9, GPQA Diamond at 84.3, LiveCodeBench v6 at 85.9, SWE-Bench Verified at 66.4, and strong German-language numbers across the same suites. Aleph Alpha’s claim is that Kolibri matches models with up to four times its active parameter count, such as Nemotron 3 Super, and it says the model sits on the Pareto frontier for quality versus serving cost, in both languages. The economics are the point when you are paying for the GPUs yourself.
The parts worth quoting
Two design choices stand out. First, the model is trained to abstain: with so-called abstention data and the company’s Merlin-Arthur protocol, it is supposed to say it does not know when the answer is not in the provided context, and the team tracks that accuracy as a feature rather than a bug. Second, the benchmarks beyond the public ones are proprietary vertical evaluations, covering automotive suppliers, semiconductors, the German public sector and aerospace, run against synthetic customer proxies, with the numbers climbing across successive training runs. That is a vendor’s own yardstick, and worth reading as such, but it is also exactly where a general model tends to fail in practice.
My read
The strategy is clear and, for once, not hand-wavy: own the middle of the market that cannot send its documents to a hosted API, and make deployment cheap enough that a mid-size institution can afford an always-on assistant. The open weights make the claims testable, which is the right price of entry. The open questions are the usual ones for a new model: independent evaluations, real-world German, and whether the tool-calling holds up outside the demos. For a regulated shop comparing a 3B-active open model against hosted alternatives, Kolibri just became a line item to price.
Source: the release post and the weights on Hugging Face.
