Kolibri-1
aleph-alpha · Sorti oct. 2026
Kolibri-1 is Aleph Alpha's English-German mixture-of-experts model with Apache 2.0 weights on Hugging Face, and a credible open-weight option for teams that must run German-language workloads on their own hardware. The catch is operations: the FP8 weights need about 78 GB of GPU memory, the pages we read name no hosted API or price, and every benchmark comes from Aleph Alpha's own harnesses. Right for public-sector and industrial teams that can self-host. Skip it for hosted-API needs or other languages.
Est-ce fait pour vous ?
Recommandé pour
- German and English assistants, document processing and RAG: the model was trained on 20T tokens of English and German text plus code
- Self-hosted, sovereign deployments for public administration, industrials and aerospace, the audience Aleph Alpha names
- Commercial use and redistribution of the weights under Apache 2.0
- Long documents: up to 1,048,576 tokens by extrapolation; native context is 262,144 and the card recommends serving at or below that
- Agent workflows: the card lists tool calling and a reasoning mode
Déconseillé pour
- Hosted API use: the pages we read name no hosted API or price; Aleph Alpha points to enterprise deployment support by contact
- Small hardware: the card gives a 78 GB FP8 footprint and a minimum of 2x A100 80 GB, 2x H100, 1x H200, 1x B200 or 1x B300
- Unsupervised, high-stakes decisions: the card says it is for human-reviewed use
- Decisions that need independent evaluations: Aleph Alpha says the benchmarks used its own harnesses
- Languages beyond English and German, which are the only two the card lists
Performances par tâche
Math reasoning
AIME 2025 English 96.9% on Aleph Alpha's harness; vendor-reported, untested by us
Coding
LiveCodeBench v6 85.9% on Aleph Alpha's harness; vendor-reported, untested by us
German-language tasks
The card gives a German overall score of 70.8 against 75.5 in English; both are vendor figures, untested by us
Tarifs
Entrée
Free (open weights)
Sortie
Free (open weights)
Contexte
262K native, 1M extrapolated; ~78 GB FP8
Tests de performance
Aucun changement de verdict pour l’instant
L’horloge tourne dès le premier jour : les changements apparaîtront ici à mesure que notre verdict évolue.
Sources
Journal de vérification
- Profil· Aucun changement
Agent automatisé