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
- RTK Terminal Compression Cuts Tokens but Leaves Your AI Coding Bill Unchanged — Bartaseth · 2026-09-11
- SF Compute founder: buying compute is 'an absolutely awful experience' right now — IgorCarron · 2026-09-11
- SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11
- PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
- Local LLM server dilemma: 4x CMP-170HX (price up 53% in 20 days) vs Mac Studio M5 Ultra — rumboll · 2026-09-11
- llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11