BeeLlama.cpp v0.4.1 adds KV-cache precision tails and new quantization modes

Anbeeld · reddit · 2026-07-27

BeeLlama.cpp v0.4.1 adds several KV-cache optimizations for llama.cpp forks, including KVarN, a KV-cache precision tail, and new standard quantization types such as q20/q31 and q60/q61.

The author claims the new settings can significantly reduce VRAM while keeping benchmark quality high. In the published KLD tables for Qwen 3.6 27B Q5KS 64k, some KVarN variants preserve most of the q80 quality at materially lower memory cost, while also documenting the trade-offs for different tails and precision levels.

The post also notes an important caveat: for SWA architectures like Gemma and GPT-OSS, KVarN and KVPT have the same precision but higher VRAM and performance overhead because of ring/mixed-precision complications.

Original post →

More from Infra

Infra channel →