FULL STORY
EdgeRock Open-Sourced: Constant KV Cache, Then a 4B Local Version
Rock Matrix open-sourced EdgeRock, tackling linear KV cache growth in long-context multimodal models, and followed up days later with EdgeRock-4B, a compact version claiming 99% memory savings that runs locally.
2026-10-27 ~ 2026-10-31 · 2 episodes · 7 posts
Episode 1 · EdgeRock Open-Sourced: Constant 72 MiB KV Cache for 256K Context (2026-10-27, 4 posts)
Rock Matrix has open-sourced EdgeRock, a multimodal model with O(1) memory for history that keeps the KV cache at a constant 72 MiB per sequence at 256K context—99.1% smaller than Qwen3.5-4B. Code and paper are available on GitHub under Apache-2.0.
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock architecture keeps KV cache at 72 MiB for 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock goes open source: constant-memory reasoning model with code, weights and paper — alifcoder · 2026-10-27
Episode 2 · Rock Matrix Open-Sources EdgeRock-4B with Constant-Memory KV Cache (2026-10-31, 3 posts)
Rock Matrix open-sourced EdgeRock-4B, a vision-language model with a fixed 72MiB KV cache that saves 99.1% of VRAM versus its Qwen3.5-4B base. It runs locally on a MacBook Pro at about 9 tokens/s, and disabling its Rock Memory Anchors crashes long-context accuracy from 94.1% to 0.4%.
- EdgeRock-4B open-sourced with fixed 72MiB KV cache, 99.1% less memory than base — testingcatalog · 2026-10-31
- EdgeRock-4B runs locally on MacBook: first token in 4.6s, ~9 tok/s on M5 Pro — testingcatalog · 2026-10-31
- EdgeRock-4B architecture: RMA anchors lift long-context accuracy from 0.4% to 94.1% — testingcatalog · 2026-10-31