Realtime-Venus puts two 9B models on Hugging Face

InclusionAI published two 9B Realtime-Venus checkpoints, one audio-visual and one audio-focused, in a public Apache-2.0 model repository. The model card lists a 40,960-token context for both, while the technical report says the system keeps foreground conversation running while its harness delegates work asynchronously. Builders can now inspect a full-duplex agent stack locally, but CUDA, FFmpeg, custom Transformers code, and a separate harness make this a research artifact—not a drop-in hosted API; live-agent memory still carries a bill.