DeepSeek releases a prototype agent harness The dsh-superagent repository describes a router-and-worker setup with separate worker sessions and tool presets. Its documentation says executor sessions run commands as the operating-system user without approval. It's a research prototype, not a managed security boundary.
takeaway: inspect the execution permissions before trying it. Use an isolated environment without personal credentials or production access, rather than treating a worker session as a sandbox. github.com/deepseek-ai/dsh-superagent-prototype
Gradium just released Phonon, a new family of lightweight text-to-speech (TTS) models designed for high-quality, real-time, on-device speech generation! Small but powerful & : 31M and 50M parameter models delivering impressive naturalness and accuracy. Multilingual ): Support for 5 languages — English, French, German, Spanish, and Portuguese. Blazing-fast inference 4 : As little as 7.5 ms to first audio and 24x faster than real time on Snapdragon 8 Elite Gen 5. Runs locally : Optimized for CPUs, Arm devices, Apple Silicon, and Snapdragon, without requiring a GPU or cloud API.
Open-source and ecosystem-ready: Phonon-31M is available under Apache 2.0 on the Hugging Face Hub, with day-zero support in Transformers!
