Reducto released r-1, a single-pass document parsing model designed to fold optical character recognition, layout detection, table extraction, formatting, and content grounding into one full-page processing step (MarkTechPost). The model reduces errors by 20 percent relative to Reducto’s prior multi-stage agentic pipeline and operates at $0.01 per page, addressing both accuracy and inference cost constraints in document processing workflows.
OpenBMB released MiniCPM5-2B, a 2.52 billion parameter dense language model with native 131,072 token context length that averages 53.9 across 34 benchmarks (MarkTechPost). The model outperforms Qwen3.5-4B, which scores 51.1, with particular strength in tool use, coding agents, and long-context retrieval, and is built for on-device deployment.