Text-AB-Stereo 3B · emotional dialogue

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

Context / input limit

Not collected

Maximum output

Not collected

Input modalities

Text, AudioSource for Input modalities (opens in a new tab)

Output modalities

AudioSource for Output modalities (opens in a new tab)

Reasoning

Not collected

Tool calling

Not collected

Structured output

Not collected

Streaming

Not collected

Batch processing

Not collected

Fine-tuning

Not collected

Knowledge cutoff

Not collected

Release date

2026-09-03Source for Release date (opens in a new tab)

Research disclosure date, not general API or weight availability.

Parameters

3BSource for Parameters (opens in a new tab)

Reported model size; a complete parameter total including every external encoder and evaluator is not inferred.

Architecture

Flow-matching diffusion transformer with mT5 text cross-attention and DAC-VAE audio latentsSource for Architecture (opens in a new tab)

Languages

EnglishSource for Languages (opens in a new tab)

Context caching

Not collected

No model-specific inference service or discounted prompt-cache contract is established for these research weights.

outputSampleRate

48 kHzSource for outputSampleRate (opens in a new tab)

Audio channels

Two channels, one speaker per channel, jointly modeledSource for Audio channels (opens in a new tab)

acousticRepresentation

128-dimensional DAC-VAE latents at 25 HzSource for acousticRepresentation (opens in a new tab)

Latent dimensions and frame rate are not API token limits or text-embedding dimensions.

textConditioning

Raw text through mT5 cross-attention, with language-ID embeddings; no forced alignment or explicit duration predictorSource for textConditioning (opens in a new tab)

Published research configuration

Emotional dialogue SFT, separately trained from the standard dialogue modelSource for Published research configuration (opens in a new tab)

adaptationData

English two-channel dialogue data with automatically extracted turn-level valence, arousal and dominance annotationsSource for adaptationData (opens in a new tab)

emotionControl

Learned projection of turn-level VAD vectors, concatenated with text embeddings for cross-attentionSource for emotionControl (opens in a new tab)

evaluatedEmotions

Angry, Happy, Neutral, SadSource for evaluatedEmotions (opens in a new tab)

Four categories used to create the evaluation dialogues, not an exhaustive production API emotion enumeration.

sampleSelection

32 ODE steps and 64 candidates; turn alignment prioritized, emotion accuracy and WER weighted equallySource for sampleSelection (opens in a new tab)

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

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

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.

Why it is included

Meta’s September 2026 paper and publication describe internally trained Text-AB systems and research evaluations, but supply no original checkpoint download, public API selector or production access contract. The paper’s CC-BY license covers the publication, not a released model-weight artifact. Open mT5, emotion recognizers and other evaluation components do not establish release of Text-AB weights.

Lifecycle dates

Sources reviewed