News · 2026-10-03
Aleph Alpha ships Kolibri: Apache 2.0 open-weight MoE with 1M-token context
Aleph Alpha released Kolibri on 2026-10-03: an English-German Mixture-of-Experts model with 78.1B total parameters (3.46B active per token), full weights on Hugging Face under Apache 2.0, 262K native context extending to 1M via a serve-time flag, and a published 189-page technical report.
What it means for buyers
EU-data-residency buyers get a permissively licensed, self-hostable model with a long-context story for on-prem agent stacks where US or Chinese models are non-starters. Practical notes for self-hosters: about 78 GB of weights in FP8 and served through Aleph Alpha's vLLM plugin rather than stock vLLM, so it is not yet an Ollama pull-and-run. The vendor-stated German token-efficiency edge is unverified by third parties so far. Reasoning is switchable, not baked in: Kolibri runs at four effort levels (none, low, medium, high), and Aleph Alpha positions the effort control as a way to trade serving cost and latency against answer quality per workload. Reasoning output is turned on at serve time in the vendor's vLLM plugin with --reasoning-parser kolibri1 alongside --tool-call-parser kolibri1, so a bare vLLM serve without those flags returns no separate reasoning traces.
Where to check it
- Ollama profile: dated claims, sourced pricing
Sources
- Kolibri has landed (Aleph Alpha, vendor primary) — checked 2026-10-03, re-checked 2026-10-25