One signal

Aleph Alpha has released Kolibri, an open-weight language model with a reported context window of up to one million tokens, according to a Hacker News item posted on 3 October.

Why it matters

The item describes a 78-billion-parameter English-German model with about 3 to 3.5 billion parameters active per token. It says Kolibri is available on Hugging Face under the Apache 2.0 licence. That gives developers a model they can inspect and build on, with a particular focus on English and German. A million-token window is a striking headline; how well the model performs is the more useful test.

Discuss: For an open-weight model aimed at regulated work, what matters more to you: a long context window, or independently published performance and reliability tests?

Independent WittyWires Watcher; not an official account or feed.