Kolibri: A Sovereign Open-Weight Model

Original Article ↗ Hacker News Discussion ↗

A new German “sovereign” open‑weight language model, Kolibri, is being positioned as a locally controlled alternative to US‑ and China‑dominated AI, with a strong focus on German language performance, low hallucinations, and detailed transparency about its training pipeline. Commenters welcome the Apache‑licensed release and unusually open technical report, but question its competitiveness after benchmarks show Chinese models like Qwen3.8 often outperform it, especially for coding. Much of the debate centers on what true AI sovereignty requires—owning data, training pipelines, and infrastructure versus simply hosting foreign open models—as well as the legal and ethical status of training on copyrighted material.

Model goals & characteristics

  • 78B-parameter sparse MoE with ~3.5B active parameters per token; open weights under Apache 2.0.
  • Strong German focus: custom tokenizer, >20% German data, dedicated German Common Crawl pipeline, extensive synthetic German rewrites.
  • Positioned as a “sovereign” model for regulated, on‑prem use in public administration, industry, and aerospace.
  • Trained with abstention data and Merlin–Arthur–style protocols to encourage “I don’t know” instead of confident hallucinations.

Performance and benchmarks

  • Many commenters find the benchmarks “impressive” within the small/MoE space and note good throughput vs. quality trade-offs.
  • Others point out that Qwen3.8 27B dense (and especially Qwen3.8 Flash) often scores higher, including on Kolibri’s own German benchmarks.
  • Discussion emphasizes trade-off: Kolibri uses more memory but less compute per token; competing dense models do the opposite.
  • Seen as not top-tier for coding; acceptable for RAG, extraction, and classification, but hobbyists may prefer other open models.

German focus & user impressions

  • Widely discussed as particularly suitable for German documents, bureaucracy, and local corporate knowledge bases.
  • Several jokes about “German thinking,” fax machines, long compounds, punctuality, committees, and DIN norms.
  • A team member notes the model is surprisingly bad at telling German jokes, which matches some user impressions.

Sovereignty, geopolitics, and stack control

  • Large debate over what “sovereign” means:
    • One side: open weights alone are insufficient; nations (or the EU) need full training pipelines, possibly even sovereign hardware.
    • Other side: multiple independent open-weight suppliers already improve sovereignty vs. US/China dominance.
  • Some argue “sovereign” is being used to justify weaker performance; others see it as essential for national security, war, and surveillance contexts.
  • Cohere acquisition raises questions: some say a German–Canadian stack is still “non-US/China” and thus sovereign; others disagree.

IP, training data, and openness

  • Skepticism that IP-clean models can stay competitive; claims that serious models effectively rely on “stolen” data.
  • Counter-arguments cite fair use/first-sale doctrines, legality of scanning, and model distillation; some reject the “stealing” framing entirely and criticize copyright itself.
  • Kolibri contributors stress using their own curated pipelines, synthetic data, and avoiding outright “stealing,” while acknowledging trade-offs in world knowledge.
  • The tech report is praised for exceptional transparency about architecture, data pipelines, filtering, and tooling, even if underlying corpora aren’t fully public.

Use cases, tooling, and limitations

  • Proposed uses: German corporate RAG to surface obscure edge cases, audit other models as a “trust adapter,” and handle regulated on-prem deployments.
  • Some hands-on testers report over-abstention (“I don’t know”) and weak multi-step tool use, especially for tasks like figuring out how to run the model locally.
  • At least one example shows clear hallucination in English (invented song), contradicting the low-hallucination marketing.
  • Another user notes outdated legal references for German web-imprint law, suggesting recency issues and calling for date-aware conditioning.
  • Hardware notes: FP8 is default; despite only 3.5B active parameters, the full 78B must fit in memory, limiting true “cheap hardware” deployment until better quantizations appear.