Local LLMs need 8-bit KV cache plus 4-bit weights to stay near BF16 quality
max_paperclips · x · 2026-07-24
A local-LLM inference article argues that 8-bit KV cache plus 4-bit weight quantization is the minimum viable baseline for running models locally, preserving 94-99% of BF16 quality.
- Dropping below 4-bit weights causes sharp quality loss.
- Below 8-bit KV cache, most methods are still research-grade.
- REAP only helps MoE models and often degrades quality.
- INT4 shows measurable loss on reasoning, math, and code tasks.
- The practical quality ladder is roughly FP8 > INT8 > Q5KM > Q4KM > Q2K.
More from Infra
- Synopsys CEO says AI chip design has reached an L4-like stage with 6 to 8 agents — Scobleizer · 2026-07-24
- A photonic chip test rig gets a chicken-coop heat lamp treatment — jwt0625 · 2026-07-24
- NVIDIA paper explores SOAP, Muon and other ways to scale LLM pretraining — A_K_Nain · 2026-07-24
- A Goldmine of Resources for Learning ML Systems — dhruv2038 · 2026-07-24
- DeepSeek’s compute cluster reliability role hints at a fast-changing infra org — teortaxesTex · 2026-07-24
- Open-weight models are winning by running on hardware people can actually buy — max_paperclips · 2026-07-24