Cache-to-Cache links two models via KV-cache, 2.5x faster than text
KV-cache fusion is trending on Hacker News again with installable code and pretrained weights — plus a new CLI that grades how agent-ready your site is.
Copy markdown
Two models, one shared cache
C2C fuses a source model's KV-cache straight into a target model through a trained projector, so a helper model hands over its 'thinking' without emitting a single token. The peer-reviewed result: about 2.5x lower latency than routing text between models, 3.1-5.4% higher accuracy than that text hand-off, and 6.4-14.2% over either model working alone.
Paper to drop-in, and it's trending
The technique is back on Hacker News' front page as builders find the official code and pretrained 'fuser' weights on Hugging Face — Qwen3-to-Qwen2.5 pairs ship ready, with arbitrary pairs supported via `pip install -e .`. Community wrappers are already packaging C2C as middleware for agent harnesses, so it's worth a look if your pipeline currently chains models through plain text.
Grade your site's agent experience
New this week: `npx ax audit <url>` (ora's MIT-licensed AX CLI) scores any site 0-100 and A-F on how usable it is by AI agents — discovery, docs, auth clarity, WebMCP tools — and ranks the highest-impact fixes so it slots into CI. It's part of a fast-growing 'agent experience' push that now includes Cloudflare's own Agent Readiness score. If agents are a customer, this is your Lighthouse.