---
title: dhara-chat
emoji: 🌊
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.14.0
python_version: "3.12"
app_file: app.py
pinned: false
short_description: ZeroGPU demo of dhara-250M tri-mode (AR/diffusion/self-spec)
models:
- codelion/dhara-250m
- mlx-community/dhara-250m-OptiQ-4bit
---

# dhara-chat — tri-mode (AR + block-diffusion + self-speculation)

A **ZeroGPU** demo of **dhara-250M** (`codelion/dhara-250m`): one 250M model that decodes in three modes.

- **Chat** tab — streaming chat; pick the decoding mode (AR types left-to-right, diffusion visibly unmasks `▓▒░` blocks).
- **Tri-mode compare** tab — the same prompt through AR / block-diffusion / self-speculation, with per-mode tokens/sec measured live.

Requires the `HF_TOKEN` Space secret to load the (private) model.

Stack: gradio 6.14 + transformers 5.5 + `spaces` (ZeroGPU), model in bf16 on GPU. The
GPU is attached only while a `@spaces.GPU` handler runs, then released.

> Hardware must be set to **ZeroGPU** in the Space settings (it is not a README field).