Max RevittNotes from building software, local models and agent tooling.https://maxrevitt.com/Paging a 27B model's memory to disk and backhttps://maxrevitt.com/posts/kv-cache-elision/https://maxrevitt.com/posts/kv-cache-elision/How we kept Qwen3.8-27B going for 457,000 tokens without compaction by evicting and restoring its attention cache on MLX and llama.cpp, what that did and didn't prove, and where MDBX fits in.Tue, 29 Sep 2026 00:00:00 GMT