How faxl works
Every agent turn resends the whole conversation. A prefill cache saves the model from reprocessing what it has already read. faxl is a prefill cache, built differently.
faxl is a proxy. Your agents talk to faxl, faxl talks to your models, and the saved state stays on disk you own.
On SSD, not GPU memory
Most caches live in GPU memory, which is small and expensive, so entries get evicted within minutes. faxl keeps them on SSD, measured in terabytes. It means cron-based daily agent jobs always hit a warm cache.
Survives restarts
A GPU cache is gone when the server stops. faxl's is still there after a reboot or a model switch. Use a vision model, then go back to coding, and the cache is hot for both.
Works on hybrid models
Granite, Kimi Linear, Qwen3.8 and Ling mix attention with recurrent layers. Ordinary caches can't save their state. faxl can. The newest, most efficient models are the ones you can run.
Caches the answer too
faxl saves the answer as it's delivered, so the next turn starts warm. Long conversations stay fast instead of slowing down.
Folders with quotas
Split the cache into folders, each with its own size limit. One busy project can't evict another's work.
You can see it
The console shows every request: hit or miss, tokens skipped, time to first reply. You can tune the cache from evidence rather than guesswork.