Researchers from the Hong Kong University of Science and Technology and CMU, in collaboration with Yanshan Technology, have developed and open-sourced a new model called EdgeRock-4B. This model significantly reduces KV Cache usage by 99.1%, allowing for much longer context windows without increased memory consumption. EdgeRock-4B also boasts a 6.42x increase in throughput compared to its base model, Qwen3.5-4B, while maintaining strong reasoning capabilities. The innovation lies in a novel memory mechanism called Rock Memory Anchors (RMA), which selectively stores crucial details from the context, enabling efficient long-context inference. AI
IMPACT Enables more efficient long-context inference, potentially reducing hardware costs and enabling new applications for LLMs.
RANK_REASON Open-source release of a new model with significant technical innovations in inference efficiency. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- CMU
- EdgeRock-4B
- Gemma 4 E4B
- Hong Kong University of Science and Technology
- KV cache
- LFM2.5-2.6B
- Mamba
- Micron
- MiniCPM5 2B
- Qwen3.5-4B
- rwkv
- Yanshan Technology
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
