Aleph Alpha ships Kolibri, a 78B Apache-2.0 open-weight model
A sparse MoE that fires just 3.46B active params, with a 1M-token window and strong German/English; plus Pi pod sandboxes your coding agent on your own server.
Copy markdown
Kolibri-1: 78B sparse MoE, Apache 2.0, runs lean
Aleph Alpha's new open-weight model activates only 3.46B of its 78B params per token and takes up to a 1,048,576-token context (262K recommended), tuned for German and English. It posts 75.5 (EN) / 70.8 (DE) overall on post-training evals and 96.5 on English math, and ships under Apache 2.0 — the weights are on Hugging Face now, so you can self-host a frontier-grade bilingual model today.
Pi pod: run your coding agent in your own sandboxes
A new open-source tool spins up isolated 'pod' sandboxes for pi coding-agent sessions on a server you control, with RBAC session sharing and browser/native-app surface automation. It's a weekend self-host for teams that don't want agents running loose on their main machine; a hosted option is 'coming soon.'