Nice. browsers have supported this for a long time, and it was annoying ffmpeg did not, because that meant a lot of non-browser desktop apps couldn't view them either.
I've been waiting for this feature! Whenever I've encountered an animated WebP, I've had to drop it into ezgif.com to convert it to mp4, now I can just write a quick Bash function to transform it with ffmpeg.
So far, Claude has helped find missing backports for the upcoming 9.0 release."
Although i was under the impression that they ususally preferred hand optimized assembly.
Not sure what their LLM/AI contribution policy looks like compared to other fundamental OSS projects.
> Although i was under the impression that they usually preferred hand optimized assembly.
Using LLMs/agents to do gap analysis and fill boilerplate doesn't rule out also reviewing the output and hand-optimising. That is how the tools should be used (if you aren't being a luddite like me and not using them at all) rather than click-and-hope vibe-coding.
And as a ffmpeg user with some old/weird hardware, I would much prefer click-and-hope support to no support at all. Ideally they’d have the resources for a real live human to hand-code assembly for every codec for every platform, but that’s probably not realistic. I’ll take what I can get and dust off my assembly skills if a click-and-hope implementation is close but not quite enough.
The problem I see is that LLM use deters many potential contributors. I understand that in your use case this is not an issue since you prefer working code over theoretical contributors (as said makes a lot of sense), but I am noticing this in many projects that transitioned hard into an AI dependency. It puts a barrier to some people. If 99% of a project's contributions are via AI, is that project still alive?
Yes, contribution is what keeps a project alive. It's up to the people involved to maintain quality. That's management. Two different things, both important.
In FFmpeg's twitter page it says that several of their developers got six free months of Claude Max 20x plan through Anthropic's Claude for Open Source Program, and that it was used, so far, to help find missing backports for this 9.0 release.
Though there is one part I didn’t understand in that interview. They were complaining of being overwhelmed by AI submitted bug reports (fair), including for obscure codecs that must have been used by a couple of users at most. And therefore implying that securing those codecs is low priority/important.
I don’t understand that. To me the severity has nothing to do with how popular is a code path, but whether that code path is accessible to an attacker. If I upload a specially crafted .mkv with a little known codec on YouTube and they use ffmpeg to process it, and I compromise YouTube’s infrastructure that way, it’s a pretty big deal, no matter the popularity of that codec.
That sounds nice in theory, but it seems obvious to me there is a major discrepancy between who is burdened with this responsibility, and who benefits from the result. Given many of these contributors are unpaid volunteers, maybe the infrastructure provider needs to secure FFmpeg in another way, for example by restricting codecs or by running it in a container?
The creator of borg and k8s of course already run such tools for untrusted inputs in containers, but there is also an endless stream of container escapes.
It's incredible how much it does, how quickly it does it, and how it just works every time. Whether I'm trying to deal with gigs of 4k video, embedding multiple subtitles into an mkv file, or just fiddling around with trimming a few seconds of audio off a wav, ffmpeg does it without skipping a beat.
For those interested in the work being done, with a bit more details that just the Changelog, I wrote a longer blogpost about the work here: https://jbkempf.com/blog/2026/ffmpeg-9.0/
More details about the ongoing work on Swscale rewrite, on the various Vulkan changes, the assembly detailed and a bit of statistics about this release.
I still hope a future FFmpeg release will make Intel QSV encoding available on Windows laptops where the manufacturer disabled this capability from the ACPI tables. The only way to use it on these laptops is currently FFmpeg on Linux.
I know this is hyperbole to make a point even this is ridiculous. A lot of YouTubers mess around with their audio/video files, and am sure they have heard of ffmpeg.
Only slightly broken already. Dont remember the details, but Gemini could not this to work just by reading the manual. https://github.com/timonoko/Skipperi
Seems Michael Niedermayer wrote that manually, without AI. Good for him. :P
ffmpeg is great, I think nobody disputes this. I use it in two ways mostly:
1) one, via mpv, and
2) two, as conversion tool primarily
ffmpeg also has many really powerful filters, but these are very confusing to use IMO and not elegant at all. I'd wish we could use some kind of simple meta-language or so, in part similar to virtualdub/avisynth (not necessarily suggesting the same API or DSL here, but just the main idea to think of multimedia data as tangible to manipulation as if it were an object oriented system or datastream system; every time I have to use ffmpeg's filter system, I ask myself if nobody designs any of this ...).
Not a Lex Friedman fan, but I highly recommend the recent podcast he did with two leading engineers (and founder) from the ffmpeg project: https://www.youtube.com/watch?v=nepKKz-MzFM
In a world full of endless AI slop, it is refreshing to see that there are still folks out there hand-rolling assembler code to squeeze out another 5% efficiency.
Changelog:
- Extend AMF Color Converter (vf_vpp_amf) HDR capabilities
- LCEVC track muxing support in MP4 muxer
- Playdate video encoder and muxer
- Add v360_vulkan filter
- HE-AAC 960 decoding (DAB+)
- transpose_cuda filter
- Add AMF Frame Rate Converter (vf_frc_amf) filter
- SMPTE 2094-50 metadata support and passthrough
- ProRes RAW VideoToolbox hwaccel
- APV Vulkan hwaccel
- Animated WebP decoder
- Animated WebP demuxer
- Remove CELT decoding support (doesn't affect Opus CELT)
- Remove ogg/celt parsing
- Bitstream filter to split Dolby Vision multi-layer HEVC
- Add AMF hardware memory mapping support.
- ONNX Runtime DNN backend with GPU execution provider support
- Remove deprecated NVENC options and support for pre-11.1 SDK versions
> - Animated WebP decoder
Nice. browsers have supported this for a long time, and it was annoying ffmpeg did not, because that meant a lot of non-browser desktop apps couldn't view them either.
I've been waiting for this feature! Whenever I've encountered an animated WebP, I've had to drop it into ezgif.com to convert it to mp4, now I can just write a quick Bash function to transform it with ffmpeg.
> - ONNX Runtime DNN backend with GPU execution provider support
Oh, wonder what fancy new things this will enable. Any examples in the wild already perhaps?
Some people might want to remove it, but onnx is one of the few ways you can run hw accelerated convolutions
I wonder if it’s going to be some custom encoder/decoder using DNN. I would also want to see some examples!
I'm curious if they used LLMs to do the tedious work of porting the (de/en)coders to different architectures, and how useful they are here.
Seems like Yes. https://nitter.net/FFmpeg/status/2084084810813743614 "Thank you @ClaudeDevs and @AnthropicAI for supporting FFmpeg through the Claude for Open Source program!
So far, Claude has helped find missing backports for the upcoming 9.0 release."
Although i was under the impression that they ususally preferred hand optimized assembly. Not sure what their LLM/AI contribution policy looks like compared to other fundamental OSS projects.
> Although i was under the impression that they usually preferred hand optimized assembly.
Using LLMs/agents to do gap analysis and fill boilerplate doesn't rule out also reviewing the output and hand-optimising. That is how the tools should be used (if you aren't being a luddite like me and not using them at all) rather than click-and-hope vibe-coding.
And as a ffmpeg user with some old/weird hardware, I would much prefer click-and-hope support to no support at all. Ideally they’d have the resources for a real live human to hand-code assembly for every codec for every platform, but that’s probably not realistic. I’ll take what I can get and dust off my assembly skills if a click-and-hope implementation is close but not quite enough.
Makes sense.
The problem I see is that LLM use deters many potential contributors. I understand that in your use case this is not an issue since you prefer working code over theoretical contributors (as said makes a lot of sense), but I am noticing this in many projects that transitioned hard into an AI dependency. It puts a barrier to some people. If 99% of a project's contributions are via AI, is that project still alive?
Yes, contribution is what keeps a project alive. It's up to the people involved to maintain quality. That's management. Two different things, both important.
No domain knowledge at all here but "find missing backports" doesn't sound like writing/porting code.
In FFmpeg's twitter page it says that several of their developers got six free months of Claude Max 20x plan through Anthropic's Claude for Open Source Program, and that it was used, so far, to help find missing backports for this 9.0 release.
Sounds like advertisement though.
That's how sponsorship works, it's an exchange of goods or funds for visibility of an audience. Nothing wrong with that
Sounds like getting them hooked on LLM so they come back for more.
No wonder both drug dealers and software developers call their customers "users".
I feel we’re blessed to have a project like FFmpeg. This open source tool is so important for everyone.
I recommend watching the Lex Fridman interview with Jean-Baptiste Kempf and Kieran Kunhya. Amazing work on and amazing project.
Though there is one part I didn’t understand in that interview. They were complaining of being overwhelmed by AI submitted bug reports (fair), including for obscure codecs that must have been used by a couple of users at most. And therefore implying that securing those codecs is low priority/important.
I don’t understand that. To me the severity has nothing to do with how popular is a code path, but whether that code path is accessible to an attacker. If I upload a specially crafted .mkv with a little known codec on YouTube and they use ffmpeg to process it, and I compromise YouTube’s infrastructure that way, it’s a pretty big deal, no matter the popularity of that codec.
That sounds nice in theory, but it seems obvious to me there is a major discrepancy between who is burdened with this responsibility, and who benefits from the result. Given many of these contributors are unpaid volunteers, maybe the infrastructure provider needs to secure FFmpeg in another way, for example by restricting codecs or by running it in a container?
The creator of borg and k8s of course already run such tools for untrusted inputs in containers, but there is also an endless stream of container escapes.
It's a really good interview. I recommend it too.
Link?
https://youtu.be/nepKKz-MzFM
It's incredible how much it does, how quickly it does it, and how it just works every time. Whether I'm trying to deal with gigs of 4k video, embedding multiple subtitles into an mkv file, or just fiddling around with trimming a few seconds of audio off a wav, ffmpeg does it without skipping a beat.
I think most everyone agrees with that. FFmpeg is kind of a corner stone when it comes to multimedia.
For those interested in the work being done, with a bit more details that just the Changelog, I wrote a longer blogpost about the work here: https://jbkempf.com/blog/2026/ffmpeg-9.0/
More details about the ongoing work on Swscale rewrite, on the various Vulkan changes, the assembly detailed and a bit of statistics about this release.
I still hope a future FFmpeg release will make Intel QSV encoding available on Windows laptops where the manufacturer disabled this capability from the ACPI tables. The only way to use it on these laptops is currently FFmpeg on Linux.
They can't, unfortunately.
Linux's drivers makes that happen, not ffmpeg; ffmpeg merely calls the API.
Intel's own first party drivers simply follow the ACPI tables, Linux ignores them.
how common is this with laptop manufacturers, and why would they do this? what brands are doing this so I can know to avoid them in the future.
One of my favorite pieces of software. FFmpeg has done so much for my career. Happy to see it still going strong.
Watching ffmpeg go from a niche tool for web back ends to what it is now has been incredible.
It's still very niche though, I bet not a single youtuber or OF-girl known they use it
> I bet not a single youtuber
I know this is hyperbole to make a point even this is ridiculous. A lot of YouTubers mess around with their audio/video files, and am sure they have heard of ffmpeg.
OF guys though?
Naive question... would it be possible to make a language that describes multimedia formats that could then be automatically converted into code.
Should take a look at some of the ISO specs for various MPEG standards for example.
The container formats are defined in a way that could potentially make this possible.
Sometimes the world carrying ancient god shift position in their on throne. These updates always feel like that.
Ffmpeg has patented codes and is GPLv3, what gives?
The patents for software don’t apply in their country
Finally, some real fucking technology.
Only slightly broken already. Dont remember the details, but Gemini could not this to work just by reading the manual. https://github.com/timonoko/Skipperi
Seems Michael Niedermayer wrote that manually, without AI. Good for him. :P
ffmpeg is great, I think nobody disputes this. I use it in two ways mostly:
1) one, via mpv, and 2) two, as conversion tool primarily
ffmpeg also has many really powerful filters, but these are very confusing to use IMO and not elegant at all. I'd wish we could use some kind of simple meta-language or so, in part similar to virtualdub/avisynth (not necessarily suggesting the same API or DSL here, but just the main idea to think of multimedia data as tangible to manipulation as if it were an object oriented system or datastream system; every time I have to use ffmpeg's filter system, I ask myself if nobody designs any of this ...).
twitch.tv has new HLS streamers which break ffmpeg HLS.
Do I understand right that the last version (v8) was released 8 years ago? So this breaks a long hiatus?
FFMpeg 8.0 came out August 22nd, 2025
8.1 came out 4 months ago on March 16th, 2026
Maybe you're thinking of VLC which has been stuck with a stable v3 release for almost a decade while work on v4 continue?
https://github.com/FFmpeg/FFmpeg/releases/tag/n8.0.3
Not a Lex Friedman fan, but I highly recommend the recent podcast he did with two leading engineers (and founder) from the ffmpeg project: https://www.youtube.com/watch?v=nepKKz-MzFM
In a world full of endless AI slop, it is refreshing to see that there are still folks out there hand-rolling assembler code to squeeze out another 5% efficiency.
They used Claude for this release, as stated yesterday in their twitter page.
Yes, to find missing backports. The hand rolled aspects are still there and haven't been replaced by slop.