A new 4-billion parameter open-source model named EdgeRock-4B has been released, which significantly reduces KV cache memory usage. Developed by Rock Matrix in collaboration with researchers from HKUST and CMU, this model maintains a fixed KV cache size of 72 MiB per sequence, regardless of context length, a drastic reduction compared to models like Qwen3.5-4B. EdgeRock-4B achieves this by employing a hybrid architecture that combines recurrent states, sliding-window attention, and a novel component called Rock Memory Anchors, aiming to improve reasoning capabilities in long-context scenarios. AI
IMPACT This model's memory efficiency could enable more accessible and cost-effective deployment of large context models for complex reasoning tasks.
RANK_REASON New model release from a research collaboration with significant technical innovation in memory management for LLMs. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
