BeeLlama.cpp Boosts KV Cache Quantization
Anbeeld · reddit · 2026-07-20
BeeLlama.cpp released v0.4.0, focusing on upgrades to KV cache quantization and precision control, backed by benchmarks.
Key changes include:
- Introducing KVarN (variance-normalized KV-cache quantization) for a better precision/performance balance at the same bit-width;
- Adding a KV cache precision tail that keeps recent tokens in BF16/F16 while quantizing the rest, preventing critical context from being corrupted;
- Supporting more standard KV cache types (q60/q61/q20/q21/q30/q31) for finer VRAM trade-offs;
- Tweaking DFlash's draft-max logic and adding reasoning-loop protection.
The author also published a companion article running KLD benchmarks with Qwen 3.6 27B and Gemma 4 31B to demonstrate the impact of these changes on precision and memory.
More from Infra
- MiniMax H3 runs fully on a Galaxy S26: 5-second video with audio in ~7 minutes — BidFunny4862 · 2026-10-27
- Unconventional AI acquires inference chip firm NeuReality in 1,000x efficiency bid — NaveenGRao · 2026-10-27
- Musk: robot RAM demand will dwarf everything else, and 100% of superintelligence compute goes to deep space — elonmusk · 2026-10-27
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27