Aleph Alpha has released Kolibri, a 78.1 billion-parameter English-German mixture-of-experts model that activates only 3.46 billion parameters per token and carries a 1 million-token context window (MarkTechPost). The model was trained on 768 B200 GPUs across Germany and Finland, with over 21 percent of training data in German, and the Apache 2.0 FP8 weights run on a single B200 or H200 (The Decoder). The release reflects a push toward open-weight alternatives with explicit regional infrastructure and licensing constraints.
Google Research has integrated federated learning with trusted execution environments, moving gradient computation from user devices to attested server-side TEEs (MarkTechPost). Access policies are published to Sigstore’s Rekor log with reproducibly buildable binaries, enabling external verification of differential privacy guarantees. Separately, Google will restrict free Gemini access to its smallest model, Flash-Lite, while reserving Flash and Pro tiers for paid subscribers starting in October 2026 (The Decoder).