Weclone
🚀 One-stop solution for creating your AI twin from chat history 💡 Fine-tune LLMs with your chat logs to capture your unique style, then bind to a chatbot to bring your digital self to life.
Last verified:
What is Weclone?
WeClone is a one-stop solution for creating your digital avatar (digital twin) from chat history. It covers the complete workflow from chat data export, preprocessing, model fine-tuning using LoRA, to deployment, enabling users to quickly generate their own AI agent that mimics their unique communication style.
The tool fine-tunes large language models (LLMs) based on personal chat records from platforms like WeChat, Telegram, QQ, WeCom, and Feishu. It supports text and image modal data fine-tuning, and includes a WeClone-audio module for high-quality voice cloning using voice message data with a 0.5B large model. Users can bind their digital avatar to chatbot interfaces on multiple platforms including WeChat, QQ, Telegram, WeCom, and Feishu for multi-end intelligent interaction.
WeClone is designed for individuals interested in digital cloning, digital avatars, voice cloning, LLM fine-tuning, and creating chatbots based on their personal communication patterns. It includes built-in privacy information filtering mechanisms (removing phone numbers, ID numbers, emails, URLs, IP addresses, location names, bank account numbers, cryptocurrency wallet addresses, and age information by default), supports local deployment and fine-tuning, and ensures data remains fully controllable and secure throughout the process.
The project is open-source (AGPL-3.0 license), hosted on GitHub, and is still in rapid iteration. It uses Python 100% and requires CUDA 12.6 or above for optimal performance. The default model is Qwen2.5-VL-7B-Instruct with LoRA fine-tuning, though users can also use other models supported by LLaMA Factory.
Weclone pricing
Pricing model: Freemium
WeClone is completely free and open-source under AGPL-3.0 license. There is no paid tier or subscription. Users download the code from GitHub and run it locally on their own hardware. The only costs users may incur are: (1) computational costs for their own GPU hardware for training and inference, (2) optional API costs if they enable API online inference mode for LLM judgment during data preprocessing (compatible with any OpenAI-compatible API), and (3) electricity costs for running training. The project is currently not partnered with any platform and has not issued any cryptocurrency.
Weclone pros
- Complete end-to-end workflow from data export to deployment
- Supports multiple chat platforms: WeChat, QQ, Telegram, WeCom, Feishu
- LoRA fine-tuning reduces VRAM requirements significantly
- Built-in privacy filtering removes sensitive data automatically
- Supports local deployment ensuring data stays private and controlled
- WeClone-audio module enables voice cloning from voice messages
- Image modal data fine-tuning supported for multimodal avatars
- One-click binding to multiple chatbot platforms
- Open-source with AGPL-3.0 license
- Active development with 18k GitHub stars and 1.5k forks
- Multiple GPU training support via DeepSpeed
- Web demo and API service for inference
- Customizable blocked words list for additional privacy
- QLoRA support enables training on lower VRAM (as low as 4GB)
- Uses fast uv Python package manager for dependencies
Weclone cons
- Windows environment not rigorously tested, WSL recommended
- 7B model performance is average, 14B+ recommended for better results
- Requires significant VRAM (16GB minimum for LoRA with 7B model)
- Not suitable for production environments without thorough testing
- Tool calling not supported after fine-tuning
- Cannot guarantee 100% privacy data removal by default filters
- Chat data quantity and quality heavily impact final results
- Voice cloning only available for WeChat (WeClone-audio module)
- WhatsApp, Discord, Slack data sources still under development
- No official GUI yet, command-line interface required
- Pure Windows users cannot use vllm offline inference
- Training requires substantial chat data (around 20k messages recommended)
- Model output may be inaccurate or misleading
- No memory support implemented yet
- Context could be richer with participant info and timestamps
Frequently asked questions about Weclone
What is WeClone?
WeClone is a one-stop solution for creating your digital avatar (digital twin) from chat history. It fine-tunes large language models using your personal chat records to capture your unique communication style, then binds to chatbot platforms so you can interact with your digital self. The complete workflow includes chat data export, preprocessing, model fine-tuning with LoRA, and deployment.
What chat platforms does WeClone support?
WeClone currently supports WeChat (text, images, voice), QQ, Telegram (text, images, location, links), WeCom (Enterprise WeChat), and Feishu. WhatsApp, Discord, and Slack are marked as 'coming soon' or 'under development' (🚧) for data sources. For deployment, WeClone supports WeChat personal accounts (based on openclaw-weixin), Telegram, Discord, and Slack.
What are the hardware requirements for running WeClone?
The project uses Qwen2.5-VL-7B-Instruct by default with LoRA fine-tuning, requiring approximately 16GB VRAM for 7B models. For QLoRA at 4-bit precision, you need as little as 6GB VRAM for 7B models. CUDA 12.6 or above is required. The full training requirements vary by model size and method: 7B LoRA needs 16GB, 14B LoRA needs 32GB, 30B LoRA needs 64GB, and 70B LoRA needs 160GB VRAM.
How do I prepare chat data for WeClone?
For Telegram, use Telegram Desktop to export chat history by clicking the top right corner, selecting 'Export chat history', choosing Photos for message types and JSON for format. Place exported ChatExport_* folders in ./dataset/telegram. For WeChat, use PyWxDump to extract chat records, export as CSV, and place CSV files in ./data/csv. The data will then be processed using the weclone-cli make-dataset command.
Does WeClone protect my privacy?
Yes, WeClone includes built-in privacy filtering that removes phone numbers, ID numbers, emails, URLs, IP addresses, geographic location names, international bank account numbers, cryptocurrency wallet addresses, and age information by default using Microsoft Presidio. It also provides a blocked_words list in settings.jsonc for custom filtering. The tool supports local deployment and fine-tuning, ensuring data stays completely under your control and doesn't leave your machine.
What models can I use with WeClone?
WeClone defaults to ChatGLM3-6b (Chinese version) or Qwen2.5-VL-7B-Instruct (English version) with LoRA fine-tuning. You can also use any other model supported by LLaMA Factory. The project supports various fine-tuning methods including Full, Freeze, LoRA, GaLore, APOLLO, BAdam, and QLoRA at different precision levels (8-bit, 4-bit, 2-bit).
Can I use WeClone for commercial purposes?
WeClone is provided 'as is' for learning, research, and experimental purposes only. Using it for commercial purposes or providing external services requires you to bear all risks yourself. The project strongly recommends thorough testing, security assessment, and establishing content review mechanisms before production use. The authors are not responsible for any losses from production environment use including service interruption, data security issues, or legal liability.
How do I deploy my trained model to a chatbot?
You can deploy using AstrBot (recommended multi-platform framework supporting QQ, Telegram, WeChat, WeCom, Feishu, Discord, Slack) or LangBot. First execute 'weclone-cli server' to start the API service. In AstrBot, add a new service provider with OpenAI type, set API Base URL (e.g., http://172.17.0.1:8005/v1 for Docker), model as gpt-3.5-turbo, and any API Key. Remember to turn off tool calling with '/tool off_all' as fine-tuned models don't support tool calling.
Does WeClone support voice cloning?
Yes, WeClone includes a WeClone-audio module that enables high-quality voice cloning using WeChat voice messages with a 0.5B large model. This feature is specifically designed for WeChat voice message cloning and can be fine-tuned to WeChat bots. The audio module has separate dependency installation requirements that don't include xcodec.
What is the current development status of WeClone?
WeClone is still in rapid iteration phase, and current performance does not represent final results. The project has 18k GitHub stars, 1.5k forks, 12 releases (v0.3.0 latest from Jan 4, 2026), and 11 contributors. The roadmap includes: support for more data sources, richer context (participant info, time), memory support, improved multimodal support, data augmentation, GUI support, and Chain-of-Thought thinking support. The project is written 100% in Python and uses AGPL-3.0 license.