NVIDIA paper enables cross-model KV cache transfer, skipping prefill 2.7-25x faster than recompute
blaizedsouza · x · 2026-09-09
A NVIDIA research paper introduces cross-model KV cache transfer: when swapping between models in a family (routing, cascading, mid-conversation switching), the receiver reuses the source model's KV cache and skips prefill entirely, with conversion running 2.7-25x faster than reprocessing the context.
Key findings:
- LLM APIs are stateless; every turn resends the full conversation and bills it as input. Prompt caching only works on the model that produced the cache, since keys/values depend on that model's weights.
- Cross-model KV shows strong linear structure: on Qwen3 14B→32B, a single source layer explains 56% of variance in the target's keys and 32% in values, rising to 79%/65% with multiple source layers.
- The method is a closed-form ridge mapper operating per head in three steps: select top-k most predictive source layers per target layer, concatenate their KV as input, strip RoPE from keys before mapping.
Since cache hits are billed at roughly 10% of base input rate, making cache reuse portable across models could extend the biggest cost lever in LLM serving to routing and model cascades.
More from Infra
- Broadcom CEO claims open-weight models burn $100B compute for $30B revenue — zephyr_z9 · 2026-09-09
- Stealth startup Kepler Computing emerges to tackle memory shortage with 3D stacking — nordicinst · 2026-09-09
- Palantir Names Nebius Its Preferred Sovereign AI Infrastructure Partner — pdamodaran · 2026-09-09
- GLM-5.3-Flash Q4 Hits 37.4 t/s at 300k Context on M3 Ultra via Custom Kernels — IngeniousIdiocy · 2026-09-09
- Cybercabs with Starlink modules enter mass production at Giga Texas, hinting at compute-energy bull case — NinaDSchick · 2026-09-09
- Running Qwen3.8-Flash-Next on 96GB VRAM: llama.cpp settings hit 15 t/s at 130k ctx — HlddenDreck · 2026-09-09