Aleph Alpha open-sources Kolibri, a sovereign German MoE model
A European model lab just put real open weights on the table, open-source maintainers are building robots to keep machine-written code out, and the labs' CEOs are arguing about theology. Three stories this hour.
Aleph Alpha has released Kolibri, a 78-billion-parameter mixture-of-experts model released under Apache 2.0 and built deliberately around German. The German lab's new flagship activates only 3.46 billion of its parameters per token — about 4.4% of the model — and was trained from scratch in Germany and Finland on 768 B200s across roughly 24 trillion tokens, with German making up more than a fifth of the mix (over 2 trillion German tokens the team curated from the web itself plus synthetic data). It ships with a 1 million-token context window, reasoning and tool-calling modes, FP8 weights at around 78 GB, and a 189-page tech report that documents the whole pipeline — data, architecture, training, post-training, evaluation — which is the point: Kolibri is pitched at public administration, industry, and aerospace buyers who need to know exactly what is inside the model they are running. On Aleph Alpha's own (vendor-reported) evaluations it lands on the Pareto frontier of quality against serving cost in English and German, edging Qwen3.5 35B-A3B on average scores (75.5 English / 70.8 German versus 74.7 / 69.8) and clearing AIME 2025 at 96.9 against 89.6 — but it is notably weak on agentic coding, where its SWE-bench Verified score of 66.4 trails the same competitor's 73.8. The honest read: this is a sovereign-availability play, not a frontier bid, and the benchmarks are the vendor's own until someone reruns them. The Hacker News thread hit 408 points within hours, which tells you there is a real audience for open weights that aren't Anglophone-first.
System76 is banning AI-generated contributions across its Pop!_OS and COSMIC repositories — and enforcing the ban with a bot. The contribution guidelines now state that "content generated by any LLM (also known as AI) may not be contributed in issues or PRs to Pop!_OS projects including code, comments, and descriptions," and every COSMIC pull request template carries a checkbox declaring the submission LLM-free; a GitHub Action closes any pull request that ships without it. Founder Jeremy Soller tightened the policy a day later, deleting a carve-out that had allowed LLM use for research and testing because the examples "may be interpreted as condoning those usages." The stated reason is maintainer load: Soller says LLM-generated first-time contributor PRs were unplanned, rarely accepted, and failed basic testing or duplicated existing work. One carve-out remains — the cosmic-flatpak repo, where only upstream manifests are imported. It is the second major Linux distribution to draw a hard line after Debian's months-long fight over the same question, and the sharpest version of it: we argued in August that a blanket ban punishes honesty rather than bad code — The Take — Debian's LLM ban would punish honesty, not bad code — but System76 skipped the debate and shipped an enforcement workflow instead.
Sam Altman says treating AI models as a religious force is "a real safety issue." The OpenAI CEO posted on X that he is "very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models," and the framing is a veiled shot at Anthropic, whose co-founder Christopher Olah recently met religious leaders and floated confession-style practices for models, according to the New York Times — days before Pope Leo XIV remarked that "algorithms lack the spark of humanity." The irony isn't lost on anyone: Altman himself used the phrase "magic intelligence in the sky" to describe OpenAI's goal back in 2023 and 2024, and has repeatedly said he wants AI that feels like the movie Her. What to watch: whether this becomes a real fault line between the labs' public-safety pitches, or just another week of executives talking past each other while the alignment work continues underneath.
Is a German-first open model the right bet for European sovereignty, or is a language niche a dead end? Tell us in the comments.
Sources: Pop!_OS contribution guidelines
