WeClone: Turning Telegram Chat Logs into a Fine-Tuned AI Twin

What WeClone Actually Solves

Most fine-tuning projects assume you already have a clean dataset. WeClone targets the messy reality of personal chat logs. The README describes a one-stop solution that goes from raw chat history to a deployed digital avatar. That means it handles data export, preprocessing, model training, and deployment as a single flow. The intended user is someone who wants a bot that mimics their own texting style, not a generic assistant. The project explicitly supports Telegram as a data source and as a deployment target. WhatsApp, Discord, and Slack are listed as under construction, so as of the latest release, Telegram is the only fully supported path. The core value is removing the glue code between chat export and a fine-tuned model, which is normally a fragmented process involving separate tools for parsing, cleaning, and training.

The Pipeline: From Telegram Chat to LoRA Adapter

The mechanism is built around LLaMA Factory, an external training framework. WeClone defaults to the Qwen2.5-VL-7B-Instruct model and uses LoRA for parameter-efficient fine-tuning. The README states that other models and methods supported by LLaMA Factory can be used instead. The data flow starts with exporting chat history from Telegram. The tool then preprocesses that data, filtering for privacy and converting supported formats. Images are supported in the training data, but voice, video, links, and files are not. Animated emojis are converted to plain emoji. After preprocessing, the training step produces a LoRA adapter. Finally, the adapter is bound to a chatbot deployment, with Telegram as the working platform. The project also mentions support for WeChat personal accounts based on openclaw-weixin, and Discord and Slack as deployment targets, though the support table marks those as under construction. The key architectural point is that training and inference configurations live in a single file, settings.jsonc, which unifies the two stages.

Getting It Running: Commands and Configuration

Setup follows a standard Python project pattern, but with a recommended package manager. The README instructs users to install uv first. Then you clone the repository, create a Python 3.12 virtual environment, and install dependencies with uv pip install --group main -e . . After that, you copy a template configuration file: cp examples/tg.template.jsonc settings.jsonc. All training and inference settings go into this file. The README warns that CUDA version 12.6 or above is required. There is also a note that Windows has not been rigorously tested, and WSL is suggested as a runtime environment. The setup process is not a one-click installer. It assumes familiarity with command-line tools, Python environments, and GPU driver management. The dependency on uv is a convenience, but it is an extra tool you must install before anything else.

Hardware Reality Check: VRAM Tables and Model Size

The README includes a detailed VRAM table that is honest about the cost of fine-tuning. For a 7B model, full fine-tuning in bf16 requires 60GB, while LoRA in 16-bit precision needs 16GB. QLoRA drops that to 10GB at 8-bit and 6GB at 4-bit. The project defaults to Qwen2.5-VL-7B-Instruct, but the README says the performance of 7B models is average and that 14B or larger models tend to deliver better results. A 14B model with LoRA needs 32GB, and QLoRA at 4-bit needs 12GB. This is a critical constraint. The tool is not designed for CPU-only environments or consumer laptops without discrete GPUs. The table also shows that larger models scale linearly, so a 70B model is out of reach for most individuals. The bottom line is that WeClone is usable on a single high-end consumer GPU, but only if you accept 7B quality or step up to 14B with quantized training.

Data Modality Gaps and Format Limitations

The data source support table reveals clear boundaries. Telegram supports text and images, but not voice, video, links, or files. Location data is supported, and forwards are supported. Quoted messages are not. Animated emojis are downgraded to static emoji. These limitations matter because a digital avatar that cannot handle the full range of your chat history will feel incomplete. The README does not explain how image data is incorporated into training, only that it is supported since June 2025. It also does not specify the maximum chat length or how the tool handles very large exports. The quality of the resulting avatar depends heavily on the quantity and quality of your chat data, as the README states. If your conversations are mostly voice notes or shared links, WeClone will not capture that part of your style. This is a genuine gap for many users, especially those who use Telegram for media-heavy exchanges.

Licensing and Maintenance Considerations

WeClone is released under AGPL-3.0. That license has network copyleft implications: if you modify the code and offer it as a network service, you may be required to release your modifications under the same license. For a personal digital avatar project, this is usually irrelevant, but it matters if you plan to build a commercial service on top of it. The project is in rapid iteration, as the README warns, and the recent release history confirms that: v0.3.03 came out in January 2026, following v0.3.02 in August 2025 and v0.3.01 in July 2025. That cadence means you should expect breaking changes between releases. The unified settings file helps, but it does not shield you from upstream changes in LLaMA Factory or model defaults. There is no mention of a migration guide or long-term support policy. If you adopt WeClone, plan to track releases and re-test your pipeline after each update.

The Only Real Alternative: LLaMA Factory Directly

WeClone is not a novel training method. It wraps LLaMA Factory, which is itself a general-purpose fine-tuning tool. The README explicitly says you can use other models and methods supported by LLaMA Factory. So the practical alternative is to skip WeClone and use LLaMA Factory directly. The difference is that LLaMA Factory gives you full control over dataset formats, training hyperparameters, and evaluation, but it does not provide chat export or a ready-made Telegram bot. You would have to write your own data preprocessing scripts and deployment glue. WeClone bundles those steps, which is its main advantage. But if your chat data comes from a platform other than Telegram, or if you need to handle voice or video, LLaMA Factory's flexible dataset interface is a better starting point. You will spend more time on plumbing, but you will not be limited by WeClone's platform-specific assumptions.

Editorial conclusion

WeClone is for individuals who already have substantial chat history, a CUDA-capable GPU with at least 10GB VRAM, and the patience to iterate on fine-tuning. It is not for users who want a quick chatbot, expect Windows-native reliability, or need multi-platform data import today (WhatsApp, Discord, Slack are still marked as under construction). Before adopting it, verify that your chat export includes the modalities you care about, confirm your GPU meets the QLoRA 4-bit requirement of roughly 6GB for a 7B model, and review the AGPL-3.0 license if you plan any commercial or closed-source deployment. The project is in rapid iteration, so check the latest release notes for breaking changes. If your goal is a polished production assistant rather than a personal style clone, a general fine-tuning platform like LLaMA Factory offers more control, but WeClone is the only tool in this material that bundles data export, training, and a Telegram bot into one pipeline.

Official sources

1. License: AGPL-3.0
2. Project website
3. README
4. Releases
5. xming521/WeClone on GitHub