EdgeRock architecture keeps KV cache at 72 MiB for 256K context, 99.1% smaller than Qwen3.5-4B
alifcoder · x · 2026-10-27
Rock Matrix's EdgeRock architecture tackles the classic problem of KV cache growing linearly with context length.
- At 256K context, per-sequence attention KV cache stays fixed at 72 MiB, vs roughly 8 GiB for Qwen3.5-4B — a 99.1% reduction
- The core mechanism, Rock Memory Anchors (RMA), stores learned memories as directly readable KV records in a fixed-size bank, remaining accessible even after content leaves the local attention window
- The design offers O(N) generation with O(1) persistent history, which the team says is especially useful for longer reasoning tasks
An accompanying chart shows the contrast: EdgeRock's KV memory stays flat while Qwen3.5-4B's climbs to 8 GiB.
Related event: EdgeRock Open-Sourced: Constant 72 MiB KV Cache for 256K Context(4 posts)→
More from Models
- Fable 5.5 Day: Community Invited to Build on New Release — iruletheworldmo · 2026-10-27
- inclusionAI's Ling-3.1-flash-1019: 560B MoE, 25B active, 1M context, AA score up to 44 — jacek2023 · 2026-10-27
- User shows Gemini ignoring explicit custom instructions and starting actions unbidden — loriofficialita · 2026-10-27
- Dev Who Called Kimi Subscription a Scam Gets Locked Out Over False Paid-Post Flag — Scobleizer · 2026-10-27
- Blogger reverses stance: Anthropic now 'crushes' OpenAI models, opposite of 3 months ago — Scobleizer · 2026-10-27
- Power user's stress test: ChatGPT great at research, flaky at image consistency — andytechuk · 2026-10-27