---
title: Zero-GPU Side-by-Side LLM Comparison
emoji: "\U0001F916"
sdk: gradio
app_file: app.py
---

# Hugging Face Space: Zero-GPU Side-by-Side LLM Comparison

This Space compares four instruction-tuned models side-by-side in a chat interface:

- Gemma4 31B-IT
- Gemma4 26B-IT
- Qwen 3.6 35B-IT
- Qwen 3.6 27B

The UI is designed for **zero local GPU** usage. It calls Hugging Face hosted
inference endpoints via `huggingface_hub.InferenceClient`, so the Space itself only
needs CPU resources.

## What is included

- `app.py`: Gradio app with four parallel chat columns.
- `requirements.txt`: Runtime dependencies.

## Important setup

1. Add your Hugging Face token in Space Secrets:
   - `HF_API_TOKEN` (or `HUGGINGFACE_HUB_TOKEN` or `HF_TOKEN`)
2. Ensure model IDs in `MODEL_MATRIX` are available through the Hugging Face
   Inference API.
3. This project is configured to use the fixed best Unsloth checkpoints for each family:
   - `unsloth/gemma-4-31B-it-unsloth-bnb-4bit`
   - `unsloth/gemma-4-26B-A4B-it`
   - `unsloth/Qwen3.6-35B-A3B-GGUF`
   - `unsloth/Qwen3.6-27B-GGUF`

## Optional customization

- Add more precision variants in each model card inside `MODEL_MATRIX`.
- Change `MODEL_MATRIX` entries to your preferred checkpoints.
- If your model requires a different prompt format, edit `_SYSTEM_PROMPT` and
  `_build_messages` in `app.py`.

## Notes

- For private or gated models, confirm your token has access.
- If a checkpoint does not support the chat endpoint, the app tries a text-
- ZeroGPU is the `zero-a10g` hardware flavor. In Spaces config, set this at the
  Space level (the README `hardware:` field is not always respected for creation).