🤖Kolibri: 78B-Parameter English-German Model Hits Hugging Face
A 78B-parameter model for English and German lands on Hugging Face
TL;DR
Kolibri, a 78B-parameter English-German model, is now available on Hugging Face. It supports up to 1M tokens and matches models four times its size in performance. Perfect for on-premise deployments under EU regulations.
Kolibri, a 78B-parameter English-German Mixture-of-Experts Transformer, is now downloadable with full weights on Hugging Face. It supports up to 1M tokens and matches models four times its size in performance. This matters for teams needing robust, multilingual models with regulatory compliance. Kolibri Origin, a 30B total, 3B active model, was validated with a 65k token context window. The model was optimized for performance across sectors, with a focus on domain-specific language and regulatory realities. It's a game-changer for those looking to run models on-premise without sending data to third-party services.
Key Points
Kolibri: 78B total parameters, 3B active parameters, supports up to 1M tokens
Kolibri Origin: 30B total parameters, 3B active parameters, 65k token context window
Matches models four times its size in performance across math, code, long context
Optimized for performance across sectors, considering domain-specific language and regulatory realities
Available on Hugging Face under Apache 2.0 license, with full weights downloadable
Why It Matters
If you're working on multilingual projects requiring robust performance and regulatory compliance, Kolibri is a game-changer. It matches models four times its size in performance and supports up to 1M tokens. The model's on-premise deployment capabilities and regulatory compliance under EU regulations make it ideal for teams needing to avoid third-party inference services.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,567 builders reading daily.