Kempf, who led dav1d to become the dominant AV1 decoder, has published dav2d under his own domain despite AV2 not being finalized. The willingness to ship code now suggests AOMedia's two-year incubation has produced a working draft stable enough to implement against, and Kempf is positioning to repeat dav1d's dominance from an earlier starting point in the codec's lifecycle.
By submitting the post and driving it to 235 points, captain_bender amplified the signal that this is a milestone event for the codec community. The unusual upvote count for a codec post — comparable to Linus releases or new languages — reflects industry recognition that dav2d's existence validates AV2's readiness.
The editorial argues dav1d won AV1 not through algorithmic cleverness but because it was 2-3× faster than libaom, small enough for browser vendors to audit quickly, and BSD-2-Clause licensed so Mozilla, Google, and Apple could ship without legal friction. This same bar — fast, threaded, portable, unencumbered — is what dav2d must clear to repeat the win.
Jean-Baptiste Kempf — the engineer behind VLC, a co-founder of VideoLAN, and the lead on dav1d — has posted dav2d, an early decoder implementation for AV2, the Alliance for Open Media's next-generation video codec. The announcement on jbkempf.com landed on Hacker News at 235 points within hours, which for a video-codec post is the kind of signal usually reserved for Linus releases and new programming languages.
dav1d is the most widely deployed AV1 decoder in the world — Firefox, Chrome (via libgav1's competitor path on some platforms), Android, VLC, FFmpeg's `libdav1d` wrapper, and a long tail of embedded players all run it. The reason is unglamorous and engineering-shaped: dav1d shipped a permissively licensed (BSD-2-Clause), assembly-tuned, threaded decoder while the reference `libaom` decoder was still being treated as a slow research artifact. Kempf is now running the same playbook on AV2, except earlier in the codec's lifecycle than dav1d hit AV1.
AV2 itself is not finalized. AOMedia has been incubating it for roughly two years, and the bitstream is still subject to change. That dav2d exists at all — and that Kempf is willing to publish it under his own domain — is a signal that the AV2 working draft is stable enough to start writing real code against.
Video codecs are a strange corner of the industry: a decade of work by hundreds of engineers produces a spec that, on paper, anyone can implement, but in practice three or four reference implementations end up running on every screen on Earth. The decoder that wins is rarely the reference decoder. It's the one that's fast, threaded, portable, and unencumbered enough that Mozilla, Google, and Apple can ship it without a legal review meeting.
dav1d won AV1 not because it had a clever algorithm libaom didn't — it won because it was 2-3× faster on real content and small enough that browser vendors could audit it in an afternoon. That's the bar dav2d has to clear, and Kempf has spent six years learning exactly what that bar looks like.
The other reason this matters: AV2's competitive landscape is uglier than AV1's was. H.266/VVC is shipping with a working royalty pool, LCEVC is finding niche adoption, and the streaming industry is genuinely tired of paying MPEG-LA-style toll roads. AV1 adoption stalled for years on encoder economics — SVT-AV1 fixed that — but decoder ubiquity arrived faster because dav1d was already there when browsers needed it. Without an equivalent for AV2, the codec risks the fate of every "successor" that didn't have a free, fast decoder ready: VP9 vs HEVC all over again, except this time the open-source side starts with a head start.
There's also a quieter point about how open-source video infrastructure actually gets built. Kempf doesn't work at Google, Mozilla, or Netflix. VideoLAN is a small non-profit, and dav1d's funding came from a coalition of companies (including AOMedia members) who needed the work done and couldn't justify doing it themselves. That model — one stubborn maintainer, modest pooled funding, and a permissive license — has produced more deployed video code than every FAANG codec team combined. Watch whether AV2 funding consolidates around dav2d the same way.
If you ship anything that decodes video at the edge — a browser app, a native player, a WebRTC pipeline, a transcoding box — dav2d is the implementation you'll eventually link against, probably without realizing it, the same way you're using dav1d right now via Firefox or your phone's media framework. The actionable move today is small: star the repo, subscribe to release notes, and start tracking AV2 hardware decode roadmaps from Intel, AMD, Qualcomm, and Apple. Hardware AV2 decoders are 24-36 months out for consumer silicon, which means software decode (i.e., dav2d) is what your users will be running for the entire first wave of AV2 content.
If you're an encoder vendor or run a streaming service, the calculus is different. You're not going to encode AV2 in production for several years, but you'll want to start A/B testing bitrate-quality curves against AV1 and VVC as soon as a usable AV2 encoder lands (likely SVT-AV2, whenever Intel decides to ship it). The single biggest cost lever in streaming infrastructure is codec efficiency, and AV2's claimed 30%+ improvement over AV1 — if it holds — is the difference between a Netflix-scale CDN bill and a 30%-smaller one.
For everyone else, the practical takeaway is: don't bake AV1 assumptions deep into your media stack. The codec churn cycle has accelerated, and the abstraction layer between "video pipeline" and "specific codec" is exactly where most teams cut corners during the AV1 rollout. Don't do that again.
The interesting question isn't whether dav2d will be good — Kempf's track record makes that a safe bet. It's whether the AV2 bitstream will stabilize fast enough to let dav2d mature in parallel, and whether AOMedia members will fund the dav2d effort with the same pragmatism that built dav1d. If both happen, AV2 has a real shot at being the first codec to ship into browsers with a production-grade open decoder on day one — instead of three years later, the way every previous generation has gone.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.