kmswin1/moezip is the official implementation repository of MoEZip, accepted at NeurIPS 2026. MoEZip is a method that compresses the key-value cache (KV cache) in sparse mixture-of-experts (MoE) language models. The latest commit, dated October 14, 2026, touches only a single file, README.md. It fills a document that previously held nothing but a one-line title into a README containing a project description, installation instructions, and the citation format.
The README makes clear the problem this method targets. Existing KV cache compression chases out entries that look useless, but in MoE models the routing decisions toward the experts can change in that process, which may degrade generation quality. MoEZip combines Fisher information, which expresses how much the routing decision fluctuates when a single entry is removed, with an attention-based importance score to decide what to evict. The README states that, when keeping only 10% of the KV cache, it beats the state-of-the-art baseline by 24 points (2.3x) on LongBench and by 20 points (4.9x) on RULER. The implementation is built on top of NVIDIA's KVPress, so it starts with installing kvpress and running run_example.sh.
That said, what this commit changed is only 55 added lines and 1 deleted line in the README, so no code has changed. All of the figures above are values the README itself discloses, and the README names no method as the comparison target. Neither the license notice nor the required hardware specifications are presented in this commit either. In other words, the method proposal is worth reading as a document, but it is too early to judge performance from the numbers alone.