English | 中文

XTurnix Pretrained (Qwen3 0.6B)

This model is described in XTurnix: Large-Scale Self-Supervised Turn Control through Two-State Binary Decisions.

Predict turn-taking decisions from transcribed dialogue history: whether the AI should keep listening or start speaking while listening, and whether it should keep speaking or stop and listen when the user speaks while the AI is speaking.

The model maintains the AI's current state and makes a turn-taking decision whenever it receives user input:

Current AI state	XTurnix decision	XTurnix output
`listening`	The user has not finished the current turn; keep listening	`keep`
`listening`	The user has finished the current turn; start responding	`start`
`speaking`	The user input does not require the AI to yield the current turn	`keep`
`speaking`	The user input requires the AI to stop speaking and listen	`stop`

Quick Start

Transformers

from transformers import pipeline

pipe = pipeline(
    model="xcczach/xturnix-pt",
    trust_remote_code=True,
    device=0,
    dtype="auto",
)

result = pipe(
    [
        {
            "role": "user",
            "content": "Should the air purifier stay on continuously, or is it enough to run it for two hours before bed?",
        }
    ],
    state="listening",
)
print(result)

The inference pipeline automatically removes trailing punctuation from the last user message to preserve model performance.

Overlong dialogue history is truncated. The system prompt and the most recent dialogue messages are retained.

Example output:

{
    "action": "<|start|>",
    "scores": {
        "<|start|>": 0.97,
        "<|keep|>": 0.03,
    },
}

The pipeline also accepts a string directly:

result = pipe("那我们明天几点出发？", state="listening")

Batch inference:

results = pipe(
    [
        {"messages": messages_a, "state": "listening"},
        {"messages": messages_b, "state": "speaking"},
    ]
)

vLLM

Start the server:

vllm serve xcczach/xturnix-pt \
  --served-model-name xturnix \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype auto \
  --max-model-len 2048 \
  --generation-config vllm

Use the included vllm_client.py as the reference client.