Perplexity has open-sourced Lily, a Rust-based inference engine with custom Metal kernels optimized for Qwen3.6-35B-A3B on Apple Silicon. The engine achieved 1.23x prefill throughput and 1.35x decode throughput compared to MLX-LM on a 40-core, 128 GB M5 Max, according to benchmarks reported by MarkTechPost. Lily powers Perplexity’s Hybrid Compute feature and is now available to the public (MarkTechPost).
Qwen Developers open-sourced zg (zvec-grep) under Apache 2.0, a local-first search layer that unifies ripgrep, BM25, and vector search behind a single interface. The tool allows agents to move from plain-language queries to exact line spans without switching tools, according to MarkTechPost reporting (MarkTechPost). Separately, MIT Technology Review reported that enterprise deployment of agentic AI requires coordination across multiple agents, systems, and data sources while maintaining safety and workflow integration (MIT Technology Review).