Introduction

  • TL;DR: Lemon Slice announced “Lemon Slice-2,” positioning it as a 20B-parameter video diffusion transformer for real-time, interactive avatar experiences, and reported a $10.5M seed round.
  • In today’s agentic AI wave, most assistants remain text-first. Lemon Slice’s pitch is to add a video layer—interactive, streaming avatars that can be embedded via API/widgets.

Why it matters: Interactive video agents force teams to treat latency, session orchestration, and abuse prevention as first-class product requirements—not “later.”

What was announced (verifiable facts)

Funding and positioning

TechCrunch reports Lemon Slice raised $10.5M seed, with investors including Matrix Partners and Y Combinator (plus named angels), to build digital avatar tech that adds video to AI chats.

Techmeme River listed the story on 2025-12-23, reflecting broad pickup across tech news flows (no public analytics on “traffic spike,” so we avoid numeric claims).

Why it matters: Funding isn’t the story; the product claim is: “production-ready” interactive video agents require sustained compute + platform engineering.

Lemon Slice-2: model and performance claims

Architecture and throughput

On its official Research page, Lemon Slice describes Lemon Slice-2 as a video diffusion transformer + inference framework, a 20B-parameter “few-step causal model” achieving 20 FPS on a single GPU, with caching/attention optimizations and “infinite-length” video claims.

TechCrunch independently states Lemon Slice says the model is 20B parameters and can livestream at 20 FPS on a single GPU.

Why it matters: Publishing throughput/latency claims shifts the conversation from “cool demos” to measurable production constraints (GPU cost, concurrency, SLOs).

Developer integration: API + session flow (Docs)

Reference flow

Lemon Slice’s docs show a straightforward sequence:

  1. Get an API key (X-API-Key)
  2. Create a room via POST /rooms (a Daily room with an agent)
  3. Join from frontend using Daily’s JS SDK
  4. Subscribe to app-message events (bot_ready, errors)
1
2
3
4

curl -X POST https://lemonslice.com/api/rooms \
  -H "X-API-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "agent_id": "YOUR_AGENT_ID" }'

1
2
3
4
5
6
7
8

import DailyIframe from "@daily-co/daily-js";
const callObject = DailyIframe.createCallObject();
await callObject.join({ url: room_url, token });

useDailyEvent("app-message", (ev) => {
  if (ev?.data?.type === "bot_ready") console.log("Agent is ready!");
  if (ev?.data?.type === "daily_error") console.log("Error:", ev?.data?.err_msg);
});

Why it matters: The “hard part” is often session lifecycle and event handling. A concrete docs flow reduces integration risk for product teams.

Safety, privacy, and abuse prevention

TechCrunch notes Lemon Slice claims it has guardrails against unauthorized face/voice cloning and uses LLMs for moderation.

Its Privacy Policy explicitly mentions collection of pictures (potentially faces) and microphone/camera recordings with consent during interactive video calls.

Why it matters: Video avatars sit close to deepfake risk. Treat consent, retention, and policy enforcement as core engineering—alongside model performance.

Conclusion

  • Lemon Slice is betting on real-time interactive video agents, anchored by a 20B-parameter model and API/widget delivery.
  • The publicly visible integration pattern (rooms + Daily SDK + events) makes it easier to evaluate in a real app.
  • For teams, success hinges on SLOs + safety operations (consent, abuse prevention, data policy) as much as on avatar realism.

Summary

  • Lemon Slice raised $10.5M seed and launched Lemon Slice-2.
  • Lemon Slice-2 is described as a 20B-parameter video diffusion transformer with 20 FPS on a single GPU.
  • Docs show an API-driven session model built around rooms + Daily JS SDK.

#LemonSlice #DigitalAvatars #AIAgents #VideoAI #DiffusionModels #VideoGeneration #RealtimeAI #MLOps #TrustSafety #GenAI

References