RolePlay-Xing4.0-29B

This model is a fine-tuned version of Xing4.0-29B-A4B, specifically optimized for vertical domain tasks related to Roleplay. It is designed to deliver an immersive conversational experience, maintain stable character personas, and handle complex narrative progression.

Base Model Highlights

The base model is Xing4.0-29B-A4B (formerly known as the TeleChat series), developed by China Telecom Artificial Intelligence Technology Co., Ltd.

• Efficient MoE Architecture: With 29B total parameters and only 4B activated per token, it achieves an excellent balance between inference efficiency and performance.
• Extended Context Support: Natively supports a 256K context length, extensible to 512K, making it highly suitable for long-form roleplay, world-building memory, and deep multi-turn dialogues.
• Domestic Hardware Ecosystem: It is the first model of this scale trained entirely on the Ascend NPU platform with the MindSpore framework, deeply optimized for complex engineering tasks.

Training Details

• Fine-tuning Task: Roleplay / Character Persona.
• Training Data: The fine-tuning dataset comprises 6,247 high-quality multi-turn roleplay dialogues. It covers 1,346 distinct characters/personas, with an average of 6 interaction turns per dialogue. This ensures strong context retention and persona consistency during extended conversations.
• Training Method: Full-parameter fine-tuning.
• Training Hardware: Ascend NPU cluster.
• Training Framework: MindSpore.

Quickstart

Inference

roleplay-Xing4.0-29B-A4B can be accessed via an OpenAI-compatible API.

from openai import OpenAI

client = OpenAI(
    base_url="your-base-url",
    api_key="your-api-key",
)

completion = client.chat.completions.create(
    model="Xing4.0-29B-A4B",
    messages=[{"role": "user", "content": "hi! Mr Bench."}],
    temperature=1.0,
    top_p=0.95,
    extra_body={
        "repetition_penalty": 1.05,
        "skip_special_tokens": False,
        "spaces_between_special_tokens": False,
        "chat_template_kwargs": {
            "enable_thinking": True,  # Set to False to disable thinking
        },
    },
)

print(completion.choices[0].message.content)

Recommended Parameters