Compression: auxiliary.compression.provider = 'deepseek', model 'deepseek/deepseek-v4-flash' — uses DeepSeek API directly, not local model.
- Open WebUI: 4GB RAM/1 CPU limit
- SSH hardened: PasswordAuthentication no, X11Forwarding no
- Root fstab updated with noatime
- All 6 drives SMART passed, RAID0 md0 healthy