Models

IBM has released a small patch to its modular Granite Switch model, adding a better speech recognizer and decision tool without retraining the base model.

Google DeepMind has announced Carbon, its strongest Gemini model yet, with a focus on finding and fixing security flaws, and is releasing it first to cyber defenders.

Liquid AI has released a new family of small open models for text, images, audio and decisions that can be fine-tuned on a single GPU and run on devices.

Google DeepMind’s new world model generates explorable, photorealistic scenes in real time, now with native audio and characters who talk, for an early group of testers.

IBM has released a new open speech model and refreshed existing ones, aiming at more accurate transcription, translation and tracking of who said what in long recordings.

Reflection, a new US lab, has released Beam as an open model on Hugging Face, aimed at coding and agent work on teams’ own hardware.

Meta and Apple, usually rivals, worked together so Meta’s newest open model runs on Mac desktops instead of cloud servers.

Google DeepMind and Singapore’s GovTech built a small AI model that spots scam conversations on the phone itself, without sending messages anywhere.

Google DeepMind is rolling out Gemini 4 Argon more widely and explained the safeguards it added because the model is so capable in cybersecurity and biology.

Google DeepMind showed its Antigravity agent working alongside a scientist to find a promising new antibiotic candidate in a huge protein database.