KV-cache Grafting: Freezing Small Models Makes Them Stronger
Corbenci · hf · 2026-07-17
Making Frozen Small Models Cheaper and Stronger with KV-cache Grafting
This work proposes a method to boost small model capabilities and reduce inference costs without changing weights: storing verified knowledge as byte-exact KV states, which are later grafted back into new inference contexts.
Core Features
- Recovery is bit-level identical: Under fixed deterministic configurations, the logits after grafting are byte-for-byte identical to recomputed results
- The paper verified byte-exactness on 12B and 31B model scales across two GPU targets
- Provided committed input/output hashes for easy verification
Results
- On AIME 2025, frozen Gemma-4-12B improved from 80.0% to 93.3%
- On recurring cases, 8 problems originally unsolvable within a 401,026 token budget were solved using cached verified answers within 61 decode tokens
- The paper claims this equates to 6,574x fewer tokens and roughly 8,700x lower energy consumption
- Usable context length expanded from 32,768 to 2,854,766 tokens without increasing extra accelerator memory
The authors describe this as a "verified-knowledge flywheel": knowledge is verified once and reused multiple times.
Related event: KV-Cache Grafting Boosts Frozen Small Models(2 posts)→
More from Infra
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11
- LLM Serving Metrics Thread: Why TPOT and Uptime Make or Break User Experience — abhijithneil · 2026-09-11
- PlanetScale launches sharded Postgres: 768 servers acting as one, 1PB scale — dhruv2038 · 2026-09-11