🚀 RELEASES & LAUNCHES
Ugreen unveils HomeAgent local-AI hubs for smart-home control
Ugreen unveiled HomeAgent, a local-AI smart-home platform.- Homeowners get local storage, security-camera analysis and Matter-based device control; developers can extend the MA100 with Skills Hub packages.
- The HA100 and HA100 Pro run AI on-device; the HA100 has 16GB RAM and two storage bays plus a dedicated NVR bay, while the Pro adds 32GB RAM and 10GbE.
- The developer-focused MasterAgent MA100 uses Nvidia Jetson Thor T5000 hardware with up to 2,070 FP4 TFLOPS.
- Preorders opened September 3; early-bird prices are $899 for HA100, $2,999 for HA100 Pro and $9,999 for MA100 ahead of a Kickstarter campaign on October 27.
- HomeAgent keeps supported storage, footage searches, event recognition and device control on the local network instead of requiring monthly cloud storage fees.
TrueSpar open-sources Paddock Rust/C++ inference engine
TrueSpar open-sourced Paddock, a Rust/C++ LLM inference engine with custom CUDA kernels.- For developers and teams running self-hosted LLM inference with GGUF or Safetensors models.
- Paddock is MIT/Apache-2.0 licensed, ships as one binary with OpenAI- and Anthropic-style APIs, and includes custom CUDA kernels.
- On one RTX PRO 6000, Paddock reached 1,062 tok/s for Qwen3.8-27B FP8 at 32 clients with 1,024-token input/output, versus 958 for vLLM and 844 for SGLang.
- Reported benchmarks found Paddock faster than vLLM in 13 of 13 test cells and faster than llama.cpp Q8_0 by 1.5x to 37x, giving builders another open local-inference runtime.