Rock Matrix has open-sourced EdgeRock-4B, a small text-and-image reasoning model with fixed-capacity memory.
EdgeRock-4B keeps its attention KV cache fixed in size by holding the last 1,024 tokens in full and saving important older details to a separate bank.
Standard language models keep a growing record of every token they read, so long chats and long reasoning take up more GPU memory over time.
Rock Matrix reports an 82.9% mean pass@1 on AIME 2024–2026, compared to 18.8% for the strongest bounded-state model it tested.