Aleph Alpha has released Kolibri, a 78B parameter model with a 1M token context window, under the Apache 2.0 license, making its weights available on Hugging Face. The model incorporates "abstention training" to indicate when it lacks information, though its effectiveness was debated. Concurrently, discussions arose about the necessity of hard budget caps for AI services, particularly for coding agents, to prevent unexpected costs. However, implementing such caps presents technical challenges due to billing data latency and limited support from cloud providers. AI
IMPACT Provides practitioners with more open-source options and highlights the need for cost-control mechanisms in AI services.
RANK_REASON Release of an open-weights model with technical details and discussion of its training methodology. [lever_c_demoted from research: ic=1 ai=1.0]
- Aleph Alpha
- Apache Software License 2.0
- DwarfStar 4
- Hugging Face
- Kolibri
- llama.cpp
- Merl
- Redis
- Simon Willison
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →