Google released TPU Sync, an open-source C++ library that transfers KV caches and model weights across TPU HBM, host DRAM, storage, and network infrastructure (MarkTechPost). The library achieves over 90 percent lower KV transfer latency on MoE models ranging from 235B to 480B parameters, directly addressing inference bottlenecks in large-scale deployments. This represents a concrete engineering contribution to the open inference ecosystem rather than an architectural or model innovation.
Google also released Gemini 4 Hydrogen, which replaces the Flash series with a 1M-token context window, 4 levels of thinking, and pricing at $1.50 and $7.50 per million tokens (MarkTechPost). Hydrogen reports a 51 Artificial Analysis score and 57 percent performance on Terminal-Bench 4.1, and benchmarks indicate it nearly matches Claude Sonnet 5.5 while costing approximately 2.5 times more per task than GPT-6.1 Sol (The Decoder). Additionally, CopilotKit released OpenBot 1.0, an MIT-licensed, self-hostable AI coworker workspace where organizations deploy their own agents and models while all actions pass policy checks and audit trails (MarkTechPost), and llama.cpp reached v1.0.0 (Hugging Face Blog).