Mixed-precision LLM quantization
Deciding, layer by layer, which parts of a language model can drop to 4 bits and which must stay at 8.
Sole author. Three pipelines built and benchmarked; work continues.
- OPT-1.3B size reduction: 3.52× (source: Quantization experiment logs, Apr 2026)
- OPT-1.3B perplexity, mixed vs uniform 4-bit: 41.85 vs 45.19 (source: Quantization experiment logs, Apr 2026)
- Mistral-7B size: 13.8 → 5.5 GB (source: Quantization experiment logs, Apr 2026)
- Zero-shot accuracy change, 10 tasks: −5.69 pp (source: Quantization experiment logs, Apr 2026)
Not every layer is equal
Quantizing some layers barely changes the output; others break it. The job is to find out which is which before spending the bits.
Three ways to measure sensitivity
Activation error plus weight kurtosis; the Hessian trace, estimated with Hutchinson’s method; and a Hessian-guided pipeline that hands off to GPTQ.
Spend the budget
Sensitive layers keep 8 bits, the rest go to 4, under a target average. First and last layers are always protected.
Benchmark against the obvious
Every result is compared with full precision and with uniform 4-bit, on perplexity and ten zero-shot tasks.