---
title: tiyat-nano
canonical_url: "https://www.modelscope.cn/studios/Tachiwin/tiyat-nano"
md_url: "https://www.modelscope.cn/studios/Tachiwin/tiyat-nano.md"
repository: Tachiwin/tiyat-nano
last_updated: 2026-09-16
sdk_type: gradio
sdk_version: 6.17.3
downloads: 0
stars: 1
---

# tiyat-nano

> tiyat-nano - Tachiwin 在 ModelScope 创建的在线 Demo。Tachiwin's first LLM model trained from ground for text generation in all indigenous languages of Mexico.

Tachiwin/tiyat-nano 是 ModelScope 魔搭社区上的在线可交互 Demo（创空间），基于 gradio 6.17.3 构建。

- **Repository**: Tachiwin/tiyat-nano
- **SDK**: gradio
- **SDK version**: 6.17.3
- **Downloads**: 0
- **Stars**: 1
- **Last updated**: 2026-09-16

Source: https://www.modelscope.cn/studios/Tachiwin/tiyat-nano

---

# Tiyat — Text Generation Demo

Interactive demo for the Tiyat multilingual language model. Loads the latest
checkpoint from `tachiwin/tiyat_checkpoints_arch1` on startup.

## Setup

This Space needs read access to the (private) checkpoint and tokenizer repos.
Add a Space secret named `HF_TOKEN` under Settings → Repository secrets —
**do not** put the token in code, since this repo is public. Use a fresh,
scoped (read-only) token, not the one that's been pasted into the training
notebooks so far.

## Notes

- Runs on CPU by default. If you switch the Space to GPU hardware, swap
  `jax`/`jaxlib` in `requirements.txt` for the matching CUDA build
  (see https://docs.jax.dev/en/latest/installation.html) — plain `pip install jax`
  is CPU-only.
- `EXPERIMENT_BRANCH` in `app.py` selects which trained checkpoint to serve
  (`arch1` or `arch2`) — same toggle convention as the training notebook.