We did it! We built the fastest interactive avatar model Introducing LemonSlice-2.1 𝘍𝘭𝘢𝘴𝘩 ⚡ Here’s how we did it using Modal and LiveKit 👇 (note: it was not easy)

@livekit 2/ Latency is everything in conversational AI. We benchmarked ourselves against other leading avatar models. We tied for 1st at p50 and are faster at p75 → p99. For real-time conversations, p90 is often more important than the median.

@livekit /3 Methodology. To easily compare providers, we used the same @livekit agent code and swapped the avatar provider. + 50x measurements per provider + Normalized the human input by using a recording + End-to-end latency measured via audio waveforms

@livekit /4 Speed alone is not the complete story. LemonSlice is also the only provider that can do whole body, hand-held objects, cartoon mascots, and more.

5/ There was not a single silver bullet to get here. We shaved off milliseconds at every layer of the stack, including: DiT model architecture, VAE decoder optimizations, CUDA graph stuff and more. @modal has been a great partner in our quest for ultra-low latency. Learn more at Try it for yourself at

@modal @livekit Much respect for the side-by-side comparison. We don’t always get to see that kind of data.

@modal @livekit 🫡

@modal @livekit This is what great engineering 🤝 research looks like

@modal @livekit Dream team 💪

iits really good, but really leaving a lot of use cases on the table by not allowing agent creation by api, let me pass photo, voice, and knowledge and create the agent from api, use it programmatically through your hosted pipeline.. right now if we want to use it with your room pipeline we have to manually create each agent, esentially makes these avatars to be landing page marketing agents or customer support.. the api control would open this up to other use cases, such us allowing our customers to create agent avatars for their agents. if you are not ready for the scale, then i understand. it would drastically increase demand. gotta be ready for it.

Appreciate the feedback. We generally recommend using the self-managed pipeline for developers who want this type of control. If you still want to use Daily on the frontend, you can try our Pipecat integration. (And if you pair with Pipecat Cloud, you still don't have to self-host).

@modal @livekit Every milliseconds counts! We spent the past few weeks hunting down every millisecond we could find. This game is won in inches, not yards.

@modal @livekit well said!

@modal @livekit Flash speed avatars… we’re getting scary close to real-time humans

@modal @livekit Agreed! Video generation used to be the slowest part of the avatar stack. Now its one of the fastest!

@modal @livekit this feels way faster than most avatar demos I’ve seen

@modal @livekit 🏎️

@modal @livekit Se siente muy rápido 👀 buen trabajo

@modal @livekit gracias!

@modal @livekit latency looks solid here, curious how it holds in real apps

@modal @livekit latency is additive across the entire stack (LLM, TTS, etc) but our added avatar latency is now minimal

@modal @livekit This is seriously impressive. Real-time avatars are the future and you’re clearly ahead of the curve.

@modal @livekit TY! 🙌

@modal @livekit

@modal @livekit Super fast 😎

@modal @livekit Lfg

@modal @livekit appreciate u!

@modal @livekit Damn… that looks fast

@modal @livekit 🏎️🏎️🏎️

@modal @livekit Much respect for a fair side-by-side comparison. We don’t always get to see that kind of data.

@modal @livekit The data doesn't lie.

@modal @livekit Awesomeness is here and fast

@modal @livekit 🙌

@modal @livekit ⚡️⚡️⚡️

@modal @livekit We got it done 💪

@modal @livekit hell yeah

@modal @livekit always be shipping 🫡

@modal @livekit This is the first AI that didn’t feel like a tool, speed changes everything.

@modal @livekit It's all about speed with interactive avatars.

@modal @livekit This is a big improvement honestly

@modal @livekit so happy to hear it! we agree

@modal @livekit this video is an AI avatar? if so, wow.

@modal @livekit Nope, that's the real Lina 🙈

@modal @livekit Super impressive work 👏 focusing on p90+ latency is exactly what real-time systems need Love the depth across the stack optimization

@modal @livekit Thank you! The engineering and ML teams really delivered on this one 💪

@modal @livekit Can devs already use this?

@modal @livekit Yep!

@modal @livekit Amazing work pushing technology forward fast.

@modal @livekit Thank you - it was a team effort 💪

@modal @livekit Team work makes the dream work. Can we connect over a DM?

@modal @livekit Real data comparisons are always appreciated and very convincing for developers like me.

@modal @livekit For a team of engineers and research scientists, it's the only way!

@modal @livekit did you make this video with lemon slice?

@modal @livekit This Lina is real 🙈

@modal @livekit Epic - congrats!! Excited to try it out ⚡️

@modal @livekit Let us know what you think!

@modal @livekit This team doesn’t stop shipping

@modal @livekit and we're just getting started! 🚢

@modal @livekit This feels insanely fast. LemonSlice 2.1 Flash is on another level.

@modal @livekit A fast response makes all the difference.

@modal @livekit Nice! when will this be available?

@modal @livekit Developers on our Enterprise plan can start building with it today :)

@Krishnasagrawal @modal @livekit Can the real time model act like it’s doing heinous and violent things to me like pretend to skin me alive or smth ?

@Krishnasagrawal @modal @livekit You can connect a LemonSlice avatar to your own LLM with our daily or livekit integration. So... yes?

@Krishnasagrawal @modal @livekit Thank you !

@modal @livekit Shaving milliseconds off every layer is truly dedicated engineering from the whole team.

@modal @livekit Engineering + Machine Learning + Infrastructure. Truly a team effort!

@modal @livekit Very cool! How fast is this in a real app?

@frasereflex @modal @livekit Try it for yourself on our homepage :) The data doesn't lie.

@modal @livekit Nice!

@modal @livekit 🚀

@modal @livekit Love your product!

@modal @livekit TY 🫶

@modal @livekit Congrats! Does 471ms include TTS or only speech to avatar?

@modal @livekit It does not include TTS. That's the time it takes for us to return video once we get the audio from TTS.