KVPress hits 3.5× cache compression

This is the kind of infra win that matters more than flashy model releases: KV cache is often the real wall for long-context serving, and even small quality drift can be worth it if the memory savings are real.

• –A 3.5× cache cut can translate into longer contexts, higher concurrency, or lower VRAM requirements on the same hardware.
• –+0.012 PPL is impressively small, but perplexity alone does not prove retrieval quality, instruction-following, or long-context stability.
• –No retraining lowers adoption friction; if it lands cleanly in KVPress, it could be easier to slot into existing Transformers-based inference stacks.
• –The Reddit discussion already shows the right skepticism: users want the PR, the method details, and long-context benchmarks before calling it solved.