FFmpeg 9.0 “Lei”: Honoring the Late Chinese Developer Behind Its Legacy

FFmpeg is a foundational multimedia framework that can decode, encode, transcode, mux, demux, stream, filter, and play audio‑video, as well as handle subtitles and metadata. It is available both as command‑line tools ( ffmpeg, ffprobe, ffplay) and as libraries ( libavcodec, libavformat, libavfilter).

Version History

2000 – Project started by Fabrice Bellard.

2011 – Community split, leading to the Libav fork.

2012 – FFmpeg 1.0 released, establishing a stable release cadence.

2016 – 3.0 “Einstein” added hardware acceleration and modern format support.

2022 – 5.0 “Lorentz” introduced large‑scale API cleanup.

2025 – 8.0 “Huffman” integrated Whisper speech‑recognition filter.

2026 – 9.0 “Lei” added ONNX Runtime DNN backend and GPU Execution Provider support.

Codename Origin

Each major release receives a codename chosen by community members on the mailing list, usually honoring scientists, mathematicians, or engineers. In March 2026, Chinese developer Liu Qi proposed “Lei” to commemorate Lei Xiaohua, who died in 2016 at age 26. Lei had contributed extensive FFmpeg tutorials, API analyses, and sample code between 2013‑2016, helping Chinese beginners enter the audio‑video field. The proposal was accepted, and the 9.0 release carries his name.

Lei Xiaohua

Lei was a Ph.D. student at Communication University of China, researching audio‑video quality assessment, codecs, and streaming. Known as “Lei‑Shen” (Thunder God) in the Chinese community, he authored over 900 CSDN blog posts and maintained 30 public GitHub repositories containing players, streamers, mobile demos, and analysis tools such as VideoEye. His teaching style broke complex FFmpeg concepts into small, compilable examples, making him a key evangelist for FFmpeg in China.

FFmpeg in the AI Era

Modern AI models do not ingest container formats directly; they require decoded, resized, and possibly resampled tensors. FFmpeg serves as the preprocessing engine that transforms raw media into model‑ready data. Typical AI‑related tasks include:

Video model training – decoding videos, normalising resolution, frame‑rate, pixel format, extracting frames and discarding audio.

Speech model training – extracting audio, converting to mono, resampling to 16 kHz or other rates.

Video understanding – batch generation of key‑frames, short clips, and metadata for vision or multimodal models.

AI video generation – stitching generated images or clips into a video, adding frame‑rate, encoding, audio tracks, subtitles, and cover images.

Digital humans & dubbing – synchronising audio‑video length, mixing, resampling, muxing subtitles, and producing final output.

General AI users – compressing uploads, clipping segments, extracting audio, creating GIFs, burning subtitles, and converting formats for platform requirements.

OpenAI Whisper’s load_audio function internally calls FFmpeg to decode various audio inputs, mix them to mono, and resample to 16 kHz. FFmpeg 8.0 already bundled a Whisper audio filter that leverages whisper.cpp for speech‑to‑text output in text, SRT, or JSON. Version 9.0 further adds ONNX Runtime DNN backend and GPU Execution Provider support, enabling certain image‑processing models to run directly within FFmpeg filter chains.

Conclusion

As AI models evolve rapidly, the invisible yet essential role of FFmpeg becomes even more critical. Remembering contributors like Lei Xiaohua, who demystified complex technology and guided newcomers, is as important as the tool itself.