Skip to content
Model infoModelsSnapshot: 15 Oct 2026

LemonSlice / LemonSlice

LemonSlice 2.1 Flash

Latency-focused avatar model and inference runtime for live, audio-driven video.

Video generation
availableClosed weightsInteractive videoAudio-driven video
Official documentation (opens in a new tab)Reviewed 14 Oct 2026
7 of 18 specification fields collected
Output resolutionNot collectedAwaiting source review
Clip durationNot collectedAwaiting source review
Input billingSee ratesLemonSlice
Output billingSee ratesPublished units and conditions below

Specifications

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

Context / input limit
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Maximum output
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Context caching
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Reasoning
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Tool calling
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Structured output
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Streaming
SupportedSource for Streaming (opens in a new tab)
Batch processing
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Fine-tuning
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Parameters
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Architecture
Video diffusion transformer with incremental-inference and VAE/runtime optimizationsSource for Architecture (opens in a new tab)
Knowledge cutoff
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Languages
Not collected

No exact model-version contract has been verified; absent evidence is not an unsupported capability.

Release date
2026-04-17Source for Release date (opens in a new tab)
Input modalities
Text, Image, AudioSource for Input modalities (opens in a new tab)

Image sets character appearance; supplied speech drives animation. CWM action controls are recorded separately from free-text prompt support.

Output modalities
VideoSource for Output modalities (opens in a new tab)

Synchronized supplied audio may be played with video; this is not a standalone TTS model.

Provider-reported latency
471 ms mean first-byte latency; 2.04 s mean full-pipeline response in provider testingSource for Provider-reported latency (opens in a new tab)

Full-pipeline result includes third-party STT, LLM and TTS; it is not the avatar-only processing time or an SLA.

Generation controls
Gesture and emotion control without per-character trainingSource for Generation controls (opens in a new tab)

Access & pricing

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

LemonSlice

Provider source (opens in a new tab)
API model ID
flash
Access
POST https://lemonslice.com/api/liveai/sessions with X-API-Key, using LiveKit/Pipecat or WebSocket transport. Supply your own speech and language stack; the hosted pipeline is separately billed. Set model to flash.
Regions
Privacy policy describes US hosting; Enterprise advertises contracted residency options including EU and US. Endpoint-specific residency must be confirmed.
Billing basis
Charges run from session request until termination, including idle time; eligible burst calls carry a 2× multiplier.Source for Billing basis (opens in a new tab)
Session duration
Ultra: 2 hours; Enterprise: 24+ hoursSource for Session duration (opens in a new tab)

Plan/session limits; architecture claims of indefinitely long generation do not remove these limits.

Concurrent requests
Comparison table lists Starter 3, Creator 5, Professional 7, Scale 10 and Ultra 20; Enterprise custom.Source for Concurrent requests (opens in a new tab)

The Ultra hero separately says up to 100 and Enterprise hero 1000+. This conflict is unresolved; do not treat 100 as the guaranteed Ultra limit.

Billing and deployment scope
Monthly subscription credits plus overage; Lite has its own usage rate and Pro is custom. Hosted STT/LLM/TTS adds a separate plan-specific per-minute charge.Source for Billing and deployment scope (opens in a new tab)
Audio formats
Mono PCM16 chunks; base64 encoding over WebSocketSource for Audio formats (opens in a new tab)
Audio sample rates
16 kHz preferred; other declared sample rates are resampledSource for Audio sample rates (opens in a new tab)
Reference image limits
One reference image or existing avatar ID. Session schema recommends 368×560 source pixels and center-crops mismatches.Source for Reference image limits (opens in a new tab)

These are source framing dimensions, not verified native output resolution.

API compatibility
LiveKit, Daily/Pipecat and direct WebSocket audio input; synchronized audio/video delivered in a roomSource for API compatibility (opens in a new tab)
Aspect ratios
2x3, 9x16, 1x1Source for Aspect ratios (opens in a new tab)

API-wide enum; plan and model restrictions apply. CWM-1 Lite only supports 2x3.

Live session controls
Explicit termination stops billing; idle timeout defaults to 60 seconds and negative values disable that timeoutSource for Live session controls (opens in a new tab)

Prices retain their published currencies and billing units below.

Published billing units · original currency · observed 14 Oct 2026
CategoryUsageRateUnit
Avatar session overageUltra plan overage1$0.18minute of session time

Rate conditions

  1. Plan subscription required; charge includes idle time and runs from session request until termination. Included credits, burst multipliers and optional hosted speech/LLM charges are separate. No token-price conversion or billing-rounding assumption.
Pricing source (opens in a new tab)

Data & enterprise terms

Privacy policy permits retention of submitted media for model evaluation/improvement. Enterprise ZDR is an optional arrangement, not a default for every account. Exact retention periods and contractual exceptions remain unverified.

Read the service policy (opens in a new tab)

Caching

Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.

01 / INITIAL INPUTProcess & retain

The system processes eligible input and retains reusable state.

02 / LATER REQUESTCheck available state

Reuse depends on matching input, session state and the cache policy.

03 / CACHE HITReuse & generate

Reusing state can reduce repeated work. Any charges follow the billing rules below.

LemonSlice

flash
Not verified

Internal inference caching does not establish a customer-configurable prompt cache, TTL or discounted tariff.

How it works
Not collected

Internal inference caching does not establish a customer-configurable prompt cache, TTL or discounted tariff.

Minimum reusable input
Not collected

Internal inference caching does not establish a customer-configurable prompt cache, TTL or discounted tariff.

Lifetime & refresh
Not collected

Internal inference caching does not establish a customer-configurable prompt cache, TTL or discounted tariff.

How to enable it
Not collected

Internal inference caching does not establish a customer-configurable prompt cache, TTL or discounted tariff.

Usage fields
Not collected

Internal inference caching does not establish a customer-configurable prompt cache, TTL or discounted tariff.

Billing rules
Not collected

Internal inference caching does not establish a customer-configurable prompt cache, TTL or discounted tariff.

Limits & data boundaries
Not collected

Internal inference caching does not establish a customer-configurable prompt cache, TTL or discounted tariff.

No verified cache rate for this endpoint. An empty rate is not a zero charge.

Caching reviewed 14 Oct 2026

What the measurements say

Mean avatar time to first byte

LemonSlice · provider-reported
471ms

Measured from receipt of the first conditional audio chunk; provider controlled test, not total conversational response or SLA.

Evaluation source (opens in a new tab)

Mean complete conversational response

LemonSlice · provider-reported
2.04seconds

Provider test collected April 15, 2026: 5 calls × 10 responses, LiveKit agent hosted locally, Deepgram Nova-3 STT, Groq Qwen3-32B LLM and Cartesia Sonic Turbo TTS. User speech end to avatar response onset; includes upstream components.

Evaluation source (opens in a new tab)

Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.

The record behind the model

Why it is included

Provider identifies this as its own model or named model offering. Reviewed delivery is its proprietary hosted service; the service terms restrict source access and distribution. No public weights for this named offering were supplied in the reviewed evidence. Customer character IDs, upstream voice/LLM models and open SDKs are separate.

Classification evidence (opens in a new tab)

Lifecycle dates

Named release announcement17 Apr 2026Source (opens in a new tab)

Sources reviewed