Breaking

Rock Matrix Open-Sources EdgeRock-4B, a Small Vision-Language Reasoning Model With Fixed-Size Memory

October 31, 2026 at 08:02 EDT

Foundation Models
Open Source

Rock Matrix has released EdgeRock-4B as open source, a compact model that handles both text and images for reasoning tasks and uses a fixed-capacity memory design to keep its resource use constant.

Rock Matrix · EdgeRock-4B open-source release

A small text-and-image model that keeps its working memory fixed as context grows

EdgeRock-4B keeps its most recent 1,024 tokens in full detail and moves older information it judges important into a separate memory bank, so the active attention cache stays the same size no matter how long the input gets.

1,024

tokens in the fixed recent window, held in full detail

Fixed

size of the active KV cache, regardless of conversation or document length

Open

released as open source, so developers can inspect, run and adapt it locally

Context grows, the active cache does not

Illustrative shapes only. The article gives no measured values for context length or memory use.

Short input

Full context

Long input

Full context

Any length

Active cache, fixed at 1,024 tokens

How information moves through the model

New tokens arrive

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.