EdgeRock-4B Open-Sources Constant-Memory Inference for LLMs

Rockmatrix AI open-sourced EdgeRock-4B, a 4B multimodal model using fixed-size three-layer memory to compress history to O(1), consuming only 72MiB for 256K-token contexts. It achieves 82.9% on AIME and supports 512 concurrent long conversations on a single GPU.

2026-11-04 ~ 2026-11-04 · 3 related posts

• EdgeRock-4B: open-source 4B model keeps KV cache flat at 72MiB over 256K tokens — eyishazyer · 2026-11-04
• EdgeRock-4B hits 82.9% on AIME with fixed-size memory, runs 512 chats on one GPU — eyishazyer · 2026-11-04
• EdgeRock open-sources constant-memory reasoning: O(1) history replaces growing KV cache — eyishazyer · 2026-11-04