Event
YanCore Matrix Open-Sources EdgeRock-4B Fixed-Memory Model
1 reports · 1 sources
EdgeRock-4B, a model released and open-sourced by YanCore Matrix (a subsidiary of Yanshan Technology) in collaboration with universities including Hong Kong University of Science and Technology and Carnegie Mellon University, is now available on GitHub and Hugging Face. The model was converted from Qwen3.5-4B and uses the team's proposed Memory Rock Anchor (RMA) mechanism, which keeps the KV Cache from growing with context length. According to figures published by the team, EdgeRock-4B's attention KV is fixed at 72 MiB per sequence, 99.1% less than the source model at a 256K context; long-reasoning throughput on the same GPU reaches 6.42 times that of the source model, and the average pass@1 rate on AIME 2024–2026 is 82.9%.
Report timeline
- RockAI Matrix open-sources EdgeRock-4B: KV Cache does not grow with context, single-card throughput reaches 6.42x the original model
RockAI Matrix, under RockAI Technology, together with universities including the Hong Kong University of Science and Technology and Carnegie Mellon University, released and open-sourced EdgeRock-4B. The model is converted from Qwen3.5-4B, and its core is the Memory Rock Anchor (RMA) mechanism proposed by the team, which keeps the KV Cache from growing with context.
甲子光年 · 微信公众号 ·