Gradium released Phonon, a family of on-device text-to-speech models. The English Phonon 31m model is under Apache 2.0, takes 64.5 MB and starts speaking in about 7 ms on a laptop CPU.
OpenAI is handing the Open Responses API specification to the open-source Agentic AI Foundation. The spec is meant to give language models a shared API, and the move places it under neutral governance.
The Allen Institute for AI has released OlmoEarth v1.3, a satellite map embedding model, along with precomputed annual embeddings of the whole planet and a browser-based mapping tool. At its finest resolution, the model condenses a year of satellite observations into a numerical fingerprint for each 10-by-10-meter patch of ground.
Action State Group is releasing open-source tools that let businesses independently check records of what AI agents did, as agents start buying, selling and delegating work between companies. The set includes EvidenceBook, which keeps tamper-evident records, and Terms Packs for evaluating agent behavior.
When it published its inference engine, DeepSeek disclosed a figure no other lab shares: how much a single GPU running its AI earns per hour. Frandroid notes that an RTX 5090 rented online earns its owner about 135 times less.
The agent platform OpenClaw now requires Node versions that reject TLS chains signed with keys compromised in October's Node.js security issue, and refuses to update on older runtimes rather than breaking an install. The release also makes package updates and recovery safer on Windows and macOS.