What it does WeClone is an end-to-end pipeline that ingests your Telegram chat exports, scrubs them for PII with Microsoft Presidio (plus a manual blocklist), fine-tunes a local LLM via LoRA, and deploys the result as a chatbot through Telegram, Discord, Slack, or WeChat. It defaults to Qwen2.5-VL-7B-Instruct and wraps LLaMA Factory for the actual training.

The interesting bit The project treats your chat history as training data for style, not just facts. It supports image-modal fine-tuning too, so your digital twin can theoretically reference the visual context of your conversations. The README is refreshingly honest that 7B models are “average” and you need 14B+ for decent results.

Key highlights

• Single CLI entrypoint (weclone-cli) for dataset prep, training, inference, and serving
• LoRA/QLoRA support means you can train on a 6GB GPU with 4-bit quantization, or scale to multi-GPU with DeepSpeed
• Privacy scrubbing is built-in but explicitly not guaranteed—manual blocklist required for anything sensitive
• Deploys via AstrBot or LangBot; exposes an OpenAI-compatible API for integration
• Image data support added June 2025; Telegram export support added July 2025

Caveats

• Windows is untested; WSL recommended
• WhatsApp, Discord, and Slack data sources are all marked 🚧 (not yet implemented)
• Tool calling breaks after fine-tuning and must be manually disabled in your bot framework
• The authors are actively soliciting English-speaking test cases, suggesting the project has been primarily validated on Chinese-language data so far

Verdict Worth a weekend project if you have years of Telegram history and a GPU with 16GB+ VRAM. Skip it if you expect plug-and-play quality from a 7B model, or if your chat history is mostly WhatsApp.