| --- |
| title: README |
| emoji: "π" |
| colorFrom: yellow |
| colorTo: red |
| sdk: static |
| pinned: false |
| --- |
| |
| <p align="center"> |
| <img src="https://huggingface.co/IKUN-LLM/ikun-2.5B/resolve/main/logo.png" width="140" /> |
| </p> |
|
|
| ## Welcome to IKUN-LLM π |
|
|
| IKUN-LLM is an open-source organization dedicated to building a **complete LLM learning curriculum** β from tokenizer to deployment β using ikun meme culture as the teaching medium. |
|
|
| We believe learning AI should be fun. Our flagship model **ikun-2.5B** (26M params) is a Chinese meme-culture dialogue model fine-tuned on [MiniMind](https://github.com/jingyaogong/minimind). The "2.5B" in the name stands for "two and a half years of practice" β itself a meme. |
|
|
| π Try our apps: [ikun-2.5B Chat](https://huggingface.co/spaces/IKUN-LLM/ikun-2.5B-chat) | [π ikun-call Voice Chat](https://huggingface.co/spaces/IKUN-LLM/ikun-call) | [ιΈ‘δΉη Soundboard](https://huggingface.co/spaces/IKUN-LLM/jilehe) | [Dance Simulator](https://huggingface.co/spaces/IKUN-LLM/ikun-dance) | [Emoji Gallery](https://huggingface.co/spaces/IKUN-LLM/ikun-emoji) |
|
|
| ## Models & Data |
|
|
| - **[ikun-2.5B](https://huggingface.co/IKUN-LLM/ikun-2.5B)** β 26M parameter Chinese dialogue model (LoRA fine-tuned, Apache 2.0) |
| - **[CXK_IKUN_Dataset](https://huggingface.co/datasets/IKUN-LLM/CXK_IKUN_Dataset)** β ~200 meme-culture SFT dialogue samples |
|
|
| ## Learning Path |
|
|
| 12 open-source repos covering the full LLM stack, all on [GitHub](https://github.com/ikun-llm): |
|
|
| **Fundamentals** β ikun-basics Β· ikun-tokenizer Β· ikun-pretrain Β· **ikun-2.5B** (SFT) |
| **Alignment** β ikun-DPO Β· ikun-GRPO Β· ikun-Reason |
| **Advanced** β ikun-MoE Β· ikun-Distill Β· ikun-V (multimodal) |
| **Engineering** β ikun-deploy Β· **[ikun-call](https://github.com/ikun-llm/ikun-call)** (real-time voice chat) |
|
|
| ## NEW: ikun-call π |
|
|
| **η»ε€ε€ζη΅θ―** β Real-time voice conversation with ikun-2.5B. |
|
|
| The 26M model runs inference in <100ms. Combined with browser speech recognition and edge-tts, you get a phone-call-like experience with an AI that speaks ikun memes. |
|
|
| - Always-on listening β just talk, no button pressing |
| - Doubao-style calling UI β pure black, immersive, mobile-first |
| - 49MB model, zero API cost, runs anywhere |
|
|
| π [Try it live](https://huggingface.co/spaces/IKUN-LLM/ikun-call) | [GitHub](https://github.com/ikun-llm/ikun-call) |
|
|