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
- Cloud capacity crisis is now an AI architecture problem, argues new newsletter — DavidLinthicum · 2026-10-26
- SpaceX files $40B GPU financing commitment with Apollo, already at 1.7GW — XFreeze · 2026-10-26
- Meta plans to raise up to $95B to fund massive AI expansion — Polymarket · 2026-10-26
- Nebius CEO: Debt Markets Closed for Neoclouds Last Week, Leaning on Customer Advances — zephyr_z9 · 2026-10-26
- CoreWeave 5-year CDS spikes to ~1,100bps as market fears AI bubble starts as credit event — kimmonismus · 2026-10-26
- AI demand reshapes cloud computing: scarce capacity and long-term deals replace on-demand — DavidLinthicum · 2026-10-26