EngineeringThe same invariants. Two products.
FreeACT and InterACT share the engineering decisions that make a single GPU handle dozens of live RTSP feeds in production: one-batch YOLO12 + TensorRT inference, RAM-disk frame buffers on tmpfs, NVENC hardware clip encoding, WSL2 GPU passthrough for steady cycle times, and hardware-aware auto-tuning so the same software runs from a 4 GB laptop to a 16 GB workstation.
On this page- One model, batched once, single-GPU batch processing as core invariant
- YOLO12 + TensorRT FP16, what changes vs PyTorch
- WSL2 GPU passthrough, the WDDM-bypass story
- Frame buffers on tmpfs, outputs on NTFS, why I/O placement matters
- Zero-bandwidth incident clips, sharing the existing reader
- NVENC encoding, near-zero CPU clips at scale
- Hardware auto-tuning, same software, laptop to workstation