What it does AI-Writer generates Chinese web fiction—fantasy (玄幻) and romance (言情)—using the RWKV architecture, a recurrent alternative to GPT-2. It offers a Python CLI, a local web UI, and even a browser-based demo for phones. The project ships separate model weights for NVIDIA GPUs, AMD/Intel GPUs via DirectML, and pure CPU.

The interesting bit The README spends as much time explaining how neural networks work as it does on installation. The author walks through embeddings, derivatives, and probability distributions in plain Chinese—an earnest attempt to demystify the black box for readers who found the project through QQ groups rather than arXiv.

Key highlights

• Supports N/A/I GPU acceleration and CPU fallback, with 2 GB VRAM sufficient for the small models
• Includes both fantasy and romance fine-tunes, switchable in run.py
• Local web interface via server.py + static index.html
• Custom sampling strategy to improve quality on small models (documented on Zhihu)
• Training pipeline lives separately at RWKV-LM; community-written tutorials linked

Caveats

• The author explicitly labels everything here as outdated; newer 1.5B/3B/7B models live in RWKV-Runner
• Context window capped at 512 tokens, so coherence drops over longer passages
• Installation is brittle: exact Python 3.8.x, specific CUDA/cuDNN/torch versions, manual model placement

Verdict Worth a look if you’re studying early RWKV adoption or need a lightweight, offline Chinese text generator that runs on modest hardware. Skip it if you want modern scale—head to RWKV-Runner instead.