rockmatrix-ai/EdgeRock

EdgeRock Strong Reasoning at Constant Memory rockmatrix.ai  ·  English · 简体中文 Strong reasoning without a growing KV cache. EdgeRock is Rock Matrix's fixed-memory model family. EdgeRock-4B, the first open-source release, understands text and images and generates text for reasoning, coding, and agent workflows. Recurrent state summarizes the context, local attention handles recent information, and Rock Memory Anchors (RMA) preserve selected distant details. These three parts keep persistent history at a fixed capacity as a request grows, leaving more GPU memory for concurrent requests and making long-running generation more economical. Quick start · Key results · Models · Whitepaper · Documentation AIME is mean pass@1 across 2024–2026; KV is compared at 256K context. Serving throughput uses 32K input and 32K output on one Hopper 80GB GPU at each model's best measured batch. It counts output tokens only, with elapsed time including prefill. On-demand agents and software need fast generation, not just…