FFmpeg 9.0

I'll pull the FFmpeg 9.0 release notes so the paragraphs stick to real names and details, not invented figures.Release notes are sparse — checking the Changelog for concrete technical changes.The FFmpeg Project released FFmpeg 9.0 "Lei", about 4 months after FFmpeg 8.1. The official release notes point builders to the project Changelog and full Git history rather than a long feature essay. On Hacker News Best the announcement drew 411 points and 93 comments, with discussion centered on the n9.0 tag and RELEASE_NOTES on GitHub.

Technically, 9.0 is a packaging of concrete pipeline work, not a rewrite of the core CLI. The Changelog lists AMF color conversion with extended HDR paths, an AMF frame-rate converter filter, and AMF hardware memory mapping; CUDA transpose; a Vulkan 360-degree filter; ProRes RAW on VideoToolbox; APV Vulkan hardware acceleration; LCEVC track muxing in the MP4 muxer; a bitstream filter that splits Dolby Vision multi-layer HEVC; SMPTE 2094-50 metadata support and passthrough; HE-AAC 960 decoding for DAB+; animated WebP demux and decode; a Playdate video encoder and muxer; and an ONNX Runtime DNN backend with GPU execution providers. It also removes CELT decoding and Ogg/CELT parsing (Opus CELT is unaffected) and drops deprecated NVENC options plus support for pre-11.1 SDK versions.

For engineers who ship media stacks, this matters where work actually lands: GPU filter graphs, container fidelity, and format coverage. Teams on AMD AMF, Apple VideoToolbox, Vulkan, or CUDA get more of the pipeline on-device instead of bouncing through software scales and colorspace stages. Broadcast and OTT paths gain DAB+ audio, Dolby Vision layer handling, and 2094-50 metadata that must survive remux. Web and mobile ingest get first-class animated WebP. Anyone running DNN-based filters can point at ONNX Runtime with a GPU provider instead of a one-off inference glue layer. Removals force a rebuild check: old NVENC SDK pins and pure CELT-in-Ogg paths will fail until you update flags and test inputs.

In market terms, FFmpeg remains the default encode/decode/filter substrate under browsers, CDNs, NLE export, live transcode fleets, and countless CLIs that never name it. Hardware vendors keep feeding their APIs into it because product demos and real customer jobs both go through ffmpeg, libavcodec, and friends. Commercial alternatives still compete on support contracts and GUIs, but they rarely displace the open stack for headless scale or for formats that land first in FFmpeg. A major.0 cut that deepens AMF, Vulkan, VideoToolbox, LCEVC, Dolby Vision, and ONNX keeps that gravity: missing a release means slower GPU paths and more custom code for formats competitors already ship.

Practical next steps are mechanical. Bump the pin to 9.0, rebuild against current AMF, CUDA, Vulkan, VideoToolbox, and NVENC SDK 11.1+, and re-run golden jobs for HDR color, frame-rate conversion, ProRes RAW, APV, animated WebP, Dolby Vision multi-layer HEVC, and LCEVC-in-MP4. Drop CELT-only Ogg assets and any pre-11.1 NVENC options from scripts. Watch follow-on point releases for regressions in the new hardware filters and the ONNX backend under your real models and GPUs; those surfaces are new enough that production validation should not wait on a later minor.