xttsultravox

Mirror of fixie-ai/ultravox-v0_6-qwen-3-32b — speech-language multimodal model that ingests audio directly without a separate ASR stage.

Source attribution

• Upstream model: fixie-ai/ultravox-v0_6-qwen-3-32b (Apache-2.0 audio adapter + Qwen 3 base)
• Audio encoder: openai/whisper-large-v3-turbo
• LLM backbone: Qwen/Qwen3-32B
• Project: https://ultravox.ai
• Paper / code: https://github.com/fixie-ai/ultravox

Use

Serve with vLLM:

vllm serve amrhym/xttsultravox \
  --trust-remote-code \
  --tensor-parallel-size 4 \
  --max-model-len 4096 \
  --gpu-memory-utilization 0.83 \
  --served-model-name xttsultravox

Then send audio via POST /v1/chat/completions with messages[].content[].audio_url.