Ant Group · audio
Realtime-Venus-Audio
Realtime-Venus-Audio is an Apache-2.0-licensed 9B full-duplex interaction checkpoint with a documented 40,960-token context. It accepts audio and text and generates text or speech on a streaming timeline.
Downloadable checkpoint · Apache-2.0
Official sources reviewed 2026-10-04. This page documents access and specifications; it does not report a model benchmark.
Task and deployment contract
Download the Realtime-Venus-Audio/ subdirectory from the shared Realtime-Venus repository; it is a separate checkpoint from the other variant.
Both variants use the MiniCPM-o 4.5 / Omni-Flow backbone and BF16 weights. The 9B figure is the model-card specification; no independent benchmark is reported here.
Full-duplex mode keeps listening while speaking and handles interruptions. The Audio checkpoint does not use a visual encoder at inference; do not treat it as the Omni video model.
Text and speech output use the bundled Token2wav resources. These checkpoints require custom inference code, Python 3.10, CUDA and FFmpeg; they are not drop-in hosted transcription endpoints.
Asynchronous tool delegation requires the separate Realtime-Venus-Harness runtime. Model outputs containing a delegate request do not themselves execute an external task.
No hosted provider rate or availability guarantee is asserted. Evaluate interruption handling, response timing and task accuracy in the actual serving environment.
Billing and limits
No current provider rate for this exact configuration is recorded here. Open weights still require compute and operations; an unlisted rate does not mean a hosted API is free.
The model table documents a 40,960-token context for both checkpoints. Video memory is a retrieval mechanism, not evidence of an unlimited raw context.
Fast Pivot — Access hundreds of AI models from a single API — competitive pricing, reliable uptime, enterprise-grade infrastructure.