Official public weights · Apache-2.0
APPLIED AI / PERSONAL INTELLIGENCE
Lucian AI.
From local conversations to a voice of my own.
A personal chatbot built on official Qwen3-1.7B weights, with my own GPU inference implementation. First, make it conversational. Then, explore my own wording and expression.
Custom tokenizer and GPU forward pass
Multi-turn history · Streaming output
No personal style fine-tuning yet
01 / CONVERSATION
Talk to Lucian AI
02 / ENGINEERING
From weights to a conversation.
The official model provides the foundation. My implementation connects each layer.
Text and tokenisation
Custom byte BPE, pre-tokenisation and role templates turn text into tokens for inference.
Local inference
Read official weights and run a custom PTX GPU forward pass with causal attention and a KV cache.
Continuous conversation
Stream generated text, manage multi-turn history, and report context trimming and reply limits.
03 / PERSONAL VOICE
Give expression a personal source.
I hope it can answer questions, and gradually reflect how I express myself.
Next, curate my own writing and ideas. Explore how personal text shapes wording, explanations and rhythm, and use independent evaluation to guide each iteration.
- 01
Make it conversational
Current stageImplement local Qwen inference and multi-turn dialogue, and inspect issues in real answers.
- 02
Build a personal corpus
PlannedCurate my own text, preserve sources, and separate training candidates from evaluation samples.
- 03
Explore personal style adaptation
Research directionCompare expression and answer quality, then use the findings to guide the next iteration.
04 / PROJECT NOTES
Record the work and its limits.
Explain what works, what comes next, and where the model’s capabilities end.
Explore my other projectsWhat can I try now?
The local model supports text and multi-turn Chinese conversation. This page starts with clearly labelled, authored UI examples. Start the local model service to switch to actual Qwen replies.
Has it learned my personal style?
No personal style fine-tuning has been implemented yet. Corpus curation and style adaptation are future plans. Example copy is not used as my personal expression corpus.
Where do the model and data come from?
The base model is official Qwen3-1.7B public weights under Apache-2.0. Inference code is implemented in this project. Personal materials and experiment checkpoints stay local.
What are the capability limits?
The model handles text, without web search or actions. Answers may be wrong. The default context is 2048 tokens; old turns are removed with a notice when needed, and reply-length limits are reported.