Xing4.0-29B-A4B-NaturalChat

This model is a conversational fine-tune of XingChen-AGI/Xing4.0-29B-A4B, specialized for natural, engaging, and context-aware Chinese and English dialogue.

Instead of responding like a formal assistant in every turn, the model is tuned to behave more like an active conversation partner: concise when appropriate, expressive when needed, and attentive to tone, subtext, and conversational continuity.

Highlights

• Natural conversation — tuned for fluent, everyday Chinese and English dialogue with less formal, assistant-style phrasing.
• Context-aware responses — designed to follow conversational context, emotional cues, implied intent, callbacks, and natural topic transitions.
• Adaptive response length — aims to answer briefly when appropriate while retaining enough detail for more involved conversations.
• Multi-turn interaction — optimized for coherent, engaging dialogue across multiple turns rather than isolated question answering.
• Flexible personas — supports system prompts and character definitions for companion chat, roleplay, interactive fiction, and dialogue-driven applications.
• Thinking control — supports both direct conversational replies and thinking mode for prompts that require more deliberate reasoning.

What “Natural Chat” Means

This is primarily a behavior adaptation, rather than a fine-tune intended only to maximize benchmark scores. It aims to improve the choices a model makes during a conversation: what to respond to, how much to say, when to ask a question, and how to preserve the tone and relationship established in previous turns.

The model is tuned to:

• respond to the most meaningful part of a message instead of mechanically addressing every detail;
• vary response length according to the conversational context;
• follow mood, subtext, callbacks, and natural topic transitions;
• take conversational initiative without constantly redirecting the user;
• avoid unnecessary headings, lists, summaries, and customer-service phrasing;
• keep roleplay characters conversational instead of turning them into generic assistants with a character description;
• follow a supplied system prompt when a specific identity, tone, setting, or relationship is desired.

Base Model

Xing4.0-29B-A4B is developed by China Telecom Artificial Intelligence Technology Co., Ltd. (中电信人工智能科技有限公司). It is a Mixture-of-Experts model in the Xing series, formerly the TeleChat series, with 29B total parameters and approximately 4B parameters activated per token.

The base model uses the mHC + MLA + MTP architecture and natively supports a 256K context window, extensible to 512K with an appropriate runtime configuration.

	Xing4.0-29B-A4B-NaturalChat
Base model	XingChen-AGI/Xing4.0-29B-A4B
Total / active parameters	29B / approximately 4B
Layers	40
Hidden size	3584
Attention	MLA
Routed experts	64, with 4 active per token
Shared experts	1
Native context length	256K
Languages	Chinese and English
Modality	Text only

Training Data

The training mixture focuses on conversational behavior, including:

• natural Chinese multi-turn dialogue;
• casual and personal conversation;
• character-based interaction and roleplay;
• emotionally aware responses;
• conversational callbacks and topic transitions;
• creative dialogue and interactive-fiction scenarios;
• general instruction data used to preserve the base model's utility.

Quickstart

Because this is a custom_code model, set trust_remote_code=True.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "summergrove/Xing4.0-29B-A4B-NaturalChat"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True,
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    device_map="auto",
    dtype=torch.bfloat16,
).eval()

Chat

For ordinary conversation, disabling thinking usually produces more direct responses with lower latency.

messages = [
    {
        "role": "system",
        "content": "你是一位自然、友善的聊天伙伴，采用生活化口语表达，语气亲和自然、有温度，尝试理解用户的情绪和意图，并给予贴合语境的回应。",
    },
    {
        "role": "user",
        "content": "我本来只想躺十分钟，结果醒来天都黑了。",
    },
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)

model_inputs = tokenizer(text, return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    top_p=0.95,
    temperature=1.0,
    repetition_penalty=1.05,
    max_new_tokens=512
)

response = tokenizer.decode(generated_ids[0], skip_special_tokens=False, spaces_between_special_tokens=False)
answer = response.split("</think>")[-1].strip()
print(answer)

OpenAI-Compatible API

After deploying the model with a compatible inference server:

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="Xing4.0-29B-A4B-NaturalChat",
    messages=[
        {
            "role": "system",
            "content": "你是一位自然、友善的聊天伙伴，采用生活化口语表达，语气亲和自然、有温度，尝试理解用户的情绪和意图，并给予贴合语境的回应。",
        },
        {
            "role": "user",
            "content": "今天回家的时候突然下雨了，还好有人借了我一把伞。",
        },
    ],
    temperature=1.0,
    top_p=0.95,
    max_tokens=512,
    extra_body={
        "repetition_penalty": 1.05,
    },
)

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

Recommended Generation Parameters

The following values are suggested starting points. Different characters and applications may benefit from additional tuning.

For casual conversation, start with:

{
  "temperature": 1.0,
  "top_p": 0.95,
  "repetition_penalty": 1.05
}

Intended Use and Limitations

• Intended for: Chinese and English casual conversation, companion-style applications, character roleplay, interactive fiction, dialogue-heavy games, creative improvisation, and conversational research.
• Not intended for: high-stakes decisions, professional medical, legal, or financial advice, or autonomous real-world actions without human supervision.
• The model may generate incorrect or fabricated information. A natural tone should not be interpreted as factual reliability.
• Long conversations may still contain repetition, forgotten details, or inconsistencies despite the base model's long context window.
• Character consistency depends on the system prompt, sampling parameters, and conversation history.
• The model inherits limitations and biases from its base model and fine-tuning data. Fine-tuning may also reduce performance on tasks outside the training distribution.
• The model has not necessarily been evaluated for every language, domain, or safety-sensitive use case. Apply application-specific evaluation and safeguards before production deployment.

License

This model is released under the Apache-2.0 license, inherited from XingChen-AGI/Xing4.0-29B-A4B. Users must also comply with the terms attached to the upstream model and all datasets used during fine-tuning.

Acknowledgements

This model is based on XingChen-AGI/Xing4.0-29B-A4B, developed by China Telecom Artificial Intelligence Technology Co., Ltd.

Thanks to the XingChen team for releasing the base model and supporting the open-source community.

Citation

If you use this model, please cite both this repository and the base model:

@misc{xing4_natural_chat,
  title  = {Xing4.0-29B-A4B-NaturalChat},
  author = {TODO},
  year   = {2026},
  url    = {https://huggingface.co/summergrove/Xing4.0-29B-A4B-NaturalChat}
}

@misc{xing4_0_29b_a4b,
  title  = {Xing4.0-29B-A4B},
  author = {China Telecom Artificial Intelligence Technology Co., Ltd.},
  year   = {2026},
  url    = {https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B}
}