Stanford researchers released Memorilla, an open-source memory system that compresses document collections into 16 memory tokens for frozen language models. Testing on Qwen3-8B, Memorilla reached 84.33% accuracy on FactKG while using 108x fewer document tokens than full context approaches. (MarkTechPost) The system addresses a core constraint in production RAG workflows - balancing retrieval quality against token budget and latency without fine-tuning.
Runway released Praxis-1, an open-weight robotics model under the Apache 2.0 license, trained on video data for robot control. The model leads the World Action Model category on Nvidia’s RoboLab-120 benchmark, though the company noted it cannot yet plan complex action sequences independently. (The Decoder) Separately, Vercel Labs released the Vercel UI SDK, an Apache-2.0 beta library of 50+ UI behaviors that coding agents can compose into accessible components for React, Vue, Svelte, Solid, and Web Components. (MarkTechPost)