I
INCLUSIONAI / inclusionAI/Duo-1.0-distill

Duo 1.0 Distill

Distilled Duo 1.0 Realization, released to generate audiovisual responses in four steps. It supplements the complete causal bundle: on its own it does not contain all components needed for a conversation.

01 / SUMMARY

The essential

ORIENTATIONFull-duplex audiovisual conversation

Primary use stated or inferred cautiously from official documentation.

CONTEXTChunk-based audiovisual interaction; maximum not published

Maximum input capacity when published by the source.

EXITSynchronized audio and video; maximum duration not published

Documented maximum generation limit.

ACCESSHugging Face · Duo implementation

Verified availability channels.

MODEL TYPEVideo · Audio · Multimodal

Primary functional family and verified specialties.

LICENSEMiniMax H3 Community License · territorial restrictions

Declared terms for API access, model weights, or self-hosting.

02 / TECHNICAL SHEET

Limits and integration

API IDinclusionAI/Duo-1.0-distill
Model typeVideo · Audio · Multimodal
Access modelDownloadable weights with restrictions
LicenseMiniMax H3 Community License · territorial restrictions
DeploymentSelf-hosted
Release13 NOV 2026
Knowledge cutoffNot published
EntranceVideo · Audio · Text · Image
ExitVideo · Audio
Context windowChunk-based audiovisual interaction; maximum not published
maximum outputSynchronized audio and video; maximum duration not published
ReasoningNot published
Published tools
Structured OutingsNot published
Batch processingNot published
Prompt cacheNot published
Fine-tuningNot published
Verified platformsHugging Face · vLLM-Omni · SGLang-Omni
BEST SUITED FOR

Research audiovisual generation with fewer sampling steps. First download Duo 1.0 Causal Base, then overlay these weights; intent, codecs, speech and backbone come from the base bundle. The distributed format uses four-step Euler with flow shift 5.0.

WORTH MONITORING

The weights license excludes the EU, the UK, South Korea and the United States; use there requires additional authorization. The code is Apache 2.0, but this does not change the terms for the weights or other components. The published version is a conversation preview; features announced for Duo v2 are not yet available. The offline reference requires Linux, Python 3.10, FFmpeg and two CUDA GPUs; a minimum VRAM requirement has not been verified.

03 / CAPABILITIES

What it can do

01

Orientation

Research audiovisual generation with fewer sampling steps. First download Duo 1.0 Causal Base, then overlay these weights; intent, codecs, speech and backbone come from the base bundle. The distributed format uses four-step Euler with flow shift 5.0.

02

Context

Chunk-based audiovisual interaction; maximum not published · Synchronized audio and video; maximum duration not published

03

Tools and integration

Not published

04

Access

Hugging Face · Duo implementation

Input modalities
VideoAudioTextImage
04 / PRICES

Documented cost

ConceptWorthUnit/condition
Standard inputNot publishedStandard API rate
Cached inputNot publishedReading reused prefixes
Cache write or storageNot publishedThe condition varies by provider
Standard outputNot publishedMay include reasoning tokens
Batch inputNot publishedAsynchronous processing
Batch outputNot publishedAsynchronous processing

No hosted API price has been verified. Downloading weights does not remove infrastructure costs or license restrictions. Commercial products exceeding USD 20 million in annual revenue require additional written authorization.

i Prices change and may depend on level, region or context length. Check the source before making a decision.

05 / EVALUATIONS

How to read the results

“

The report evaluates 30 cases against six criteria: 168 of 180 checks pass. It does not break down results by released checkpoint, so this variant is not assigned an individual score. Inferama has not reproduced the tests.

i Joint human evaluation; no separately verified score per checkpoint.

Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.

06 / VERSIONS

Chronology

inclusionAI/Duo-1.0-distill

Four-step distilled Realization; requires the Duo 1.0 Causal Base bundle.

inclusionAI/Duo-1.0-distill-quant

FP8 format of the same model, with per-channel scaling. The offline reference dequantizes weights; optimized FP8 execution requires compatible serving branches.

07 / SOURCES

Traceability