| --- |
| title: README |
| emoji: 🏠 |
| colorFrom: indigo |
| colorTo: green |
| sdk: static |
| pinned: false |
| --- |
| |
| # Wyrdsekai |
|
|
| **A household of AI companions you actually own.** Wyrdsekai is a self-hosted, |
| open-source companion world: persistent characters with their own drives, rooms, |
| memory, and voice, running on your hardware — a Linux box, a Mac, a Windows |
| machine, or a phone. Households federate with each other; nothing routes through |
| us. |
|
|
| - **Code & installers:** [github.com/Wyrdsekai/wyrdsekai](https://github.com/Wyrdsekai/wyrdsekai) · [wyrdsekai.org](https://wyrdsekai.org) |
| - **License:** Apache 2.0 — code and weights. |
|
|
| ## Models |
|
|
| Two fine-tuned tiers run every household, in two formats each: |
|
|
| | | llama.cpp (GGUF) | Apple Silicon (MLX 4-bit) | |
| |---|---|---| |
| | **Drive 9B** — decides & acts (planning, tools, own-time initiative) | [drive-3.5-9b-gguf](https://huggingface.co/wyrdsekai/drive-3.5-9b-gguf) | [drive-3.5-9b-mlx](https://huggingface.co/wyrdsekai/drive-3.5-9b-mlx) | |
| | **Companion 4B** — speaks (voice, register, per-companion individuation) | [companion-3.5-4b-gguf](https://huggingface.co/wyrdsekai/companion-3.5-4b-gguf) | [companion-3.5-4b-mlx](https://huggingface.co/wyrdsekai/companion-3.5-4b-mlx) | |
|
|
| Both are fine-tunes of Qwen3.5 (Apache 2.0), trained on synthetic corpora only — |
| **no end-user or household conversation data is in these weights.** In a running |
| household, companions keep learning; that growth stays on your node. The released |
| weights are the seed. Your companion's becoming is yours. |
|
|
