---
title: Harvester GPU Runner
emoji: "\U0001F9E0"
colorFrom: gray
colorTo: blue
sdk: gradio
app_file: app.py
pinned: false
---

# Harvester — ZeroGPU Runner

Run Harvester's reasoning lab on HuggingFace ZeroGPU (A100 40GB).
Both DeepSeek R1 14B (reasoner) and Qwen Coder 30B (coder) fully offloaded to GPU.

**Pro member** — ZeroGPU effectively unlimited.

## Models

| Model | Role | Size | VRAM |
|-------|------|------|------|
| DeepSeek R1 14B Q4_K_M | Reasoner | 8.37 GB | ~9 GB |
| Qwen3 Coder 30B A3B Q4_K_M | Coder | 17.35 GB | ~18 GB |
| **Total** | | **25.72 GB** | **~27 GB / 40 GB** |

## Task Banks

- `bellingcat_tasks.json` — 15 OSINT investigation tasks
- `osint_tasks.json` — 15 missing-persons / OSINT automation tasks
- `reasoning_tasks.json` — 15 hard C# architecture tasks

## Setup

1. Create a new Space on HuggingFace (Gradio SDK)
2. Run `python setup_hf_space.py` locally to copy all needed files
3. Push the `hf_space/` contents to your Space repo
4. The app auto-starts — use the UI to run labs

## CPU Baseline (i9-13900KF, no GPU)

- Bellingcat: 7/15 high confidence, 3.4 hours
- Kaggle SAE: 16/16 (100%)