About
I'm Yuhuan, an undergraduate with a deep interest in artificial intelligence.
My work in AI runs from datasets, LLM base models, inference infrastructure and API gateways through to applications like AI agents and AI news. Outside that, I have shipped macOS applications, designed front-end websites, run Minecraft servers, and built Discord bots.
Rather than a polished demo that beats someone else's, or a product that is easy to monetise, I would rather build things that contribute to the community and applications that solve real problems.
Skills
Languages
Python・TypeScript・Swift・C/C++・SQL・Shell
Machine Learning
PyTorch・Transformers・Apple MLX
Data Engineering
Hugging Face Datasets・pandas・NumPy・PyArrow
Web
Next.js・React・Tailwind CSS・Radix UI・Framer Motion・MDX
Backend & APIs
FastAPI・SQLAlchemy・PostgreSQL・MongoDB・Redis・MinIO・WebSockets
Infrastructure
Docker・GitHub Actions・GHCR・Trivy・Ubuntu・Cloudflare・uv・pnpm
macOS Apps
Swift・SwiftUI・Core Audio・Core ML
Digital Content
Unreal Engine・Unity・DaVinci Resolve・Premiere Pro
Experience
YuhuanStudio
Founder
- ModelYunmo v0・Yunmo v0.1・Yunmo V1・Yunmo-Next
- InferenceYunshu
- DatasetWikiZH_Dataset・+17 more
- ApplicationYunmo・Yunxin・YunNEWS・YunAudio・Pocket 3 Controller・Guizhong
- WebsiteYunUI・YuhuanStudio Official
Education
YZU Electrical Engineering Program B
B.S.
The Affiliated Zhongli Senior High School of National Central University
75th
Work
- Pocket 3 ControllerA native macOS app for Pocket 3 preview, manual gimbal control and on-device AI — with CLI and MCP access built in.
- YunAudioAn audio router for macOS with its own virtual device, and a signal path that can be proved bit-exact.
- YunNEWSA daily digest of what happened in AI, in Traditional Chinese, every item carrying its original source.
- YunAgentA general-purpose agent workbench. LLM, Plan and Agent modes share one durable runtime: events, checkpoints, pause/resume/replay, MCP tools.
- YunshuA fast local LLM / VLM inference engine for Apple Silicon. One process, OpenAI- and Anthropic-compatible, running on-device via MLX — fast first token, fast lossless decode, and prefix reuse.
- YunUIA React 19 + Tailwind v4 design system that doesn't shout. Token-driven, three themes, framework-agnostic adapters — every detail quietly considered, none of it loud.
- Yunmo-NextA Chinese language model pretrained from scratch on a GQA + Kimi Delta Attention hybrid, grown layer by layer from 0.44B to 1.01B parameters with a native 32K context.
- YunxinA unified API gateway that aggregates 24+ LLM providers behind a single API supporting Chat Completions API, Responses API, and Messages API. One API key, every model.
- YunmoA digital being that keeps existing — with its own sense of time, its own emotional boundaries, and an inner life that depends on no one. Built on LinOSS and an 8-dimensional emotion vector, so that computation produces something that deserves to be called company.
- WikiZH_DatasetA Chinese Wikipedia corpus toolkit for pretraining and multimodal datasets: one parse yields plain text, interleaved image-text, and image-caption pairs.
- Yunmo V1A Traditional Chinese small language model pretrained from scratch on the minimind architecture, v0 through v1.
- AI-FriendA locally-run AI character built on LangChain, Ollama and ChatTTS, with an optional VTube Studio link to drive an avatar.
- GuizhongA feature-rich Discord bot with music, AI integration, lottery, and more.
Models & Datasets
Yunmo-Next-1B-InstructModel
The instruction-tuned Yunmo-Next-1B-Base: 1.76B packed tokens of SFT teaching a chat protocol with an optional reasoning channel, four effort levels and tool calls. A research checkpoint.
1.008B · 16K context · ~5.3 days on one RTX 4070 Ti12downloads
Yunmo-Next-1B-BaseModel
Traditional-Chinese-first hybrid pretrained from scratch: GQA → KDA → KDA → KDA cells with low-rank Attention Residuals, grown in depth from 24 to 60 layers during training.
1.008B · 60 layers · 65,536 vocab · 19.76B tokens80downloads
Yunmo-Next-TokenizerTokenizer
A 65,536-entry byte-level BPE for Traditional Chinese, English, code and maths — no UNK, lossless on any UTF-8, first on 9 of 12 domains against 18 tokenizers.
65,536 · byte-level BPE · lossless
Yunmo-V1Model
Pretrained from scratch on MiniMind v3 (Qwen3-style: RMSNorm, SwiGLU, RoPE, GQA, QK-Norm), grown from 8 to 24 layers with a 24,000-token Traditional Chinese tokenizer.
195M · 24 layers · 768 hidden · 8 heads / 4 KV · 24k vocab · RoPE 32k16downloads
Yunmo v0.1Model
Pretrained from scratch on MiniMind v2, with the vocabulary replaced by a 25,600-token Traditional Chinese tokenizer.
0.1B · 16 layers · 768 hidden · 25.6k vocab · RoPE 32kgated
Yunmo v0Model
Pretrained from scratch on MiniMind v2, keeping the upstream 6,400-token vocabulary.
0.1B · 16 layers · 768 hidden · 6.4k vocab · RoPE 2kgated
wikipedia-zh / zh-twDataset
Plain-text Chinese Wikipedia, one article per row with its section structure preserved, ready for pretraining.
Traditional + Simplified · 1.48M rows each2,185downloads15likes
wikipedia-omni-zh / zh-twDataset
Interleaved image-text Wikipedia corpus, formatted to the MMC4 and OBELICS conventions.
Traditional + Simplified · 1.48M rows each523downloads1likes
wikipedia-image-zh-twDataset
Wikipedia images, one per row, with caption, alt text and the article section they appear in.
1.83M rows110downloads1likes