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
Maximum output Not collected
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 Context caching Not collected No model-specific inference service or discounted prompt-cache contract is established for these research weights.
Output duration About one minute one-shot; long-form generation demonstrated beyond ten minutes using multi-diffusionSource for Output duration (opens in a new tab) Longer generation is an algorithmic capability, not an unlimited hosted service entitlement.
Access & pricing Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.
API model ID Not collectedDescriptive catalog identity, not a verified selectable API ID.
Access Paper describes internal speech-generation models. No public model selector, original weight download or endpoint-specific tariff is supplied.
Regions Research deployment only; public availability and residency are not established Endpoint pricing has not been collected. Check the provider for applicable rates.
Data & enterprise terms The cited research release does not supply a current production processing/retention agreement. Research access or an expired demonstration is not a production API entitlement.
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 INPUT Process & retain The system processes eligible input and retains reusable state.
02 / LATER REQUEST Check available state Reuse depends on matching input, session state and the cache policy.
03 / CACHE HIT Reuse & generate Reusing state can reduce repeated work. Any charges follow the billing rules below.
Meta · research reference Not verified No published service cache policy for this research record.
How it works Not collected No published service cache policy for this research record.
Minimum reusable input Not collected No published service cache policy for this research record.
Lifetime & refresh Not collected No published service cache policy for this research record.
How to enable it Not collected No published service cache policy for this research record.
Usage fields Not collected No published service cache policy for this research record.
Billing rules Not collected No published service cache policy for this research record.
Limits & data boundaries Not collected No published service cache policy for this research record.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
What the measurements say
Short-form dialogue human-likeness Meta research authors · provider-reported 4.53MOS / 5 Table 4, held-out conversations averaging about 30 seconds; real recordings score 4.62 in the same evaluation. Provider-run human evaluation.
Evaluation source (opens in a new tab)
Long-form dialogue human-likeness Meta research authors · provider-reported 3.86MOS / 5 Table 4, synthetic longer scripts versus the previous internal system’s 3.00. Raters review sampled one-minute segments; not the same protocol as the short-form comparison.
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 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.
Classification evidence (opens in a new tab) Lifecycle dates Sources reviewed Model notes Nine current/historical Muse identities plus closed-weight research models, trained variants and named pipeline components, including 2026 Text-Audiobox research. API availability is explicitly separated from research references. Regional/private models, historical successors and hosted partner routes remain under review.
Research reference only. No public API selector or model-specific price has been verified. Current Muse tariffs, token limits and processing policies do not apply to this record.
Research configuration, not a public API product. No inference price, character quota, cache discount, streaming contract or fine-tuning service is established.
Audio references are optional in the general framework. A reference-free generation can sample a voice; it does not imply a hosted inventory of preset voices.
Full-duplex here describes generated two-channel conversations, not a verified live, bidirectional voice-agent API.
Section 3.1 summarizes dubbing SFT before branching, while sections 3.3 and 5.1 specify initialization from the pretrained mono checkpoint. The training lineage ambiguity is retained rather than silently rewritten.
Check against your workload Try representative inputs, measure task success and total cost, then check the exact endpoint’s data policy.
Read our methodology
Also from Meta Muse Wonder Muse Wonder Full Muse Spark 1.3 Muse Spark 1.2 Muse Spark 1.1 Muse Spark · original April release Muse Image Muse Voice Transcribe Muse Video · preview Movie Gen Video Personalized Movie Gen Video Movie Gen Edit