One page: how to shrink AI models to run on the edge for a video product. The two levers (distillation trains a smaller student model; quantization stores the same model's numbers in fewer bits) and how they compound; the bit-width memory math for a 7-billion-parameter model (28 / 14 / 7 / 3.5 GB at FP32 / FP16 / INT8 / INT4) against an 8 GB edge ceiling; the six quantization methods (INT8 + calibration, GPTQ, AWQ, GGUF Q4_K_M, FP8, NVFP4) and what each is for; PTQ-first / QAT-if-needed; the per-stage decision for a video pipeline (detection on INT8/TensorRT, speech on a distilled-then-quantized model, language/VLM on 4-bit AWQ/GGUF); and the questions to ask before you compress, including re-measuring accuracy on your own footage.
Download free PDF
The book · Volume 1 of 7
Video Encoding: A Complete Guide to Video Compression — From Pixels and Color to H.264, HEVC, AV1, Quality Metrics, and AI
This checklist shows how distillation and quantization shrink AI models to fit an 8 GB edge device, with memory math and a per-stage plan for a video pipeline. Volume 1 covers the video side in depth: from pixels and color to H.264, HEVC and AV1, quality metrics for checking results on your own footage, and AI in compression. Written by Nikolay Sapunov, CEO at Fora Soft.
Need your detection and speech models compressed to run on edge hardware?
Tell us your devices and models, and we will plan distillation and quantization per stage, then measure accuracy on your own footage. 250+ video products shipped since 2005.
You get a realistic estimate, technical recommendations and next steps. Free, no obligations.
