NVIDIA has released TensorRT Model Connect in public preview under Apache 2.0, a toolchain that compiles supported Hugging Face checkpoints or local models directly to native C++ inference in two commands, eliminating intermediate ONNX export steps and producing versioned .bundle artifacts for deployment (MarkTechPost). The project targets inference engineers seeking reduced compilation friction and faster model-to-production workflows on NVIDIA hardware.
Google has open-sourced Sovereign Agent Mesh (SAM), a zero-config, zero-trust P2P overlay network that enables autonomous agents to discover and invoke each other’s MCP tools across cloud, on-premise, laptop, and edge environments without exposing credentials or infrastructure details (MarkTechPost), released under Apache 2.0. Separately, Hugging Face published a guide on multi-vector and late-interaction embedding models for retrieval-augmented generation systems, and The Decoder reported that Artificial Analysis benchmarked seven search API providers for AI agents, with GPT-5.6 Luna, Parallel, Exa, and Firecrawl ranking highest on combined quality, cost, and speed metrics (The Decoder).