# Project Huginn

> Project Huginn is a distributed GPU sharing platform for AI model training and fine-tuning. Built by SMware ApS (Startup Denmark approved), it enables users to train LLMs at 50% lower cost using a community-powered GPU pool. GPU owners can share idle compute and earn money.

## Core Platform

- [Homepage](https://www.projecthuginn.com): Main landing page with features, pricing, and GPU tier information
- [Login](https://www.projecthuginn.com/login): Sign in to access the platform as a customer or GPU owner
- [Dashboard](https://www.projecthuginn.com/dashboard): Customer dashboard with job overview, statistics, and quick actions

## Guides (Learn) — public educational content

- [Learn hub](https://www.projecthuginn.com/learn): Guides on distributed AI training, GPU pools and fine-tuning
- [What is distributed AI training](https://www.projecthuginn.com/learn/what-is-distributed-ai-training): Data vs model parallelism, FedAvg, communication-efficient training
- [How to train AI models cheaply](https://www.projecthuginn.com/learn/train-ai-models-cheaply): Idle GPUs, LoRA/QLoRA, spot capacity, distributed pools
- [LoRA vs QLoRA](https://www.projecthuginn.com/learn/lora-vs-qlora): VRAM, speed and quality trade-offs; when to use each
- [Federated learning across devices](https://www.projecthuginn.com/learn/federated-learning-every-device): Train small models across browsers, phones, CPUs and GPUs
- [GPU cloud alternatives](https://www.projecthuginn.com/learn/gpu-cloud-alternatives): Hyperscalers vs marketplaces (Vast.ai, RunPod) vs distributed pools
- [What is a GPU pool](https://www.projecthuginn.com/learn/what-is-a-gpu-pool): Decentralized compute for AI, verification, and earning

## Distributed / Federated capabilities

- Adaptive sharding: a job is split across however many GPUs are active right now.
- FedAvg aggregation: weights from each device are averaged into a global model each round.
- Capability-aware routing: heavy jobs (YOLO/CNN/LLM) go to capable GPUs; light jobs and
  small-model federated training run on any device, including browsers (WebGPU/TensorFlow.js).
- Verified contribution: a fraction of jobs are re-run on a second device and compared;
  providers are paid proportionally to verified work; cheaters lose pay and reputation.
- Computer Vision: dataset audit, quality analysis, preprocessing, and deep model training.
- Physical AI: robotics / drones / automation workloads (training, inference, simulation, reporting).

## Model Studio (Fine-Tuning)

- [Model Studio](https://www.projecthuginn.com/dashboard/studio): Browse and manage your fine-tuned AI models
- [Train New Model](https://www.projecthuginn.com/dashboard/studio/train): 4-step wizard to fine-tune a base model with your own data
- [Huginn Assistant](https://www.projecthuginn.com/dashboard/assistant): AI-powered platform guide that answers questions about training, billing, and GPU compute

### Available Base Models (Free)
- LLaMA 3.1 8B (Meta) — General-purpose, 100+ languages, 12GB VRAM minimum
- LLaMA 3.1 70B (Meta) — Enterprise-grade, large datasets, 40GB VRAM minimum
- Mistral 7B (Mistral AI) — Fast fine-tuning, excellent cost-performance ratio
- Phi-3 Mini 3.8B (Microsoft) — Small but powerful, low VRAM, fast results
- CodeLlama 7B (Meta) — Code generation and analysis (Python, JS, SQL)
- Gemma 2 9B (Google) — Open-source, multilingual, strong reasoning
- Qwen 2.5 7B (Alibaba) — Multilingual, strong math and coding

### Training Methods
- LoRA (Low-Rank Adaptation) — Fast and efficient, ~12GB VRAM, recommended for most use cases
- QLoRA (Quantized LoRA) — 4-bit quantized, ~8GB VRAM, budget-friendly
- Full Fine-Tune — All weights trained, ~40GB+ VRAM, highest quality

### Training Data Formats
- JSONL (Alpaca format): {"instruction": "...", "input": "...", "output": "..."}
- CSV: instruction,input,output columns
- Chat Format (JSONL): {"messages": [{"role": "user", "content": "..."}, ...]}
- Maximum file size: 500MB

## Jobs & Compute

- [My Jobs](https://www.projecthuginn.com/dashboard/jobs): View and monitor all submitted training and inference jobs
- [Submit New Job](https://www.projecthuginn.com/dashboard/jobs/new): Create a new compute job with profile selection and cost estimation

## Billing & Pricing

- [Billing](https://www.projecthuginn.com/dashboard/billing): Manage HU balance, view transactions, and top up credits
- 1 HU (Hugin Unit) = €0.21
- Pay-per-use model with pre-estimation and upper bound guarantees
- No monthly fees or subscriptions

### GPU Tiers & Rates
| Tier | GPUs | VRAM | HU/Hour |
|------|------|------|---------|
| T1 | H100, A100 80GB | 40GB+ | 4.50 |
| T2 | A100 40GB, RTX 4090 | 20GB+ | 2.80 |
| T3 | RTX 4080, RTX 3090 | 12GB+ | 1.60 |
| T4 | RTX 3080, RTX 2080 Ti | 8GB+ | 0.90 |
| T5 | Entry-level, Mobile | <8GB | 0.40 |

## Earn From Any Device

- Earn from virtually any device: desktop GPU, laptop, tablet, smartphone, electric vehicle, or even a browser tab
- Connect via the Huginn Agent app or open a browser tab — that's it
- Desktop GPUs earn the most (up to €0.60/hour for high-end cards)
- Smartphones, tablets, and browser tabs generate passive income 24/7
- Electric vehicle owners can earn while driving using their car's compute power
- Owner payout rate: 1 HU = €0.153 net (73% owner share)
- Automatic job scheduling and workload distribution
- Payouts via Stripe in EUR

## Company

- **Company**: SMware ApS
- **Country**: Denmark
- **Status**: Startup Denmark Approved
- **Website**: https://www.projecthuginn.com
- **Year Founded**: 2025

## Technical Details

- Platform: Next.js web application
- Database: PostgreSQL (Supabase)
- Payment: Stripe integration
- GPU Agent: Downloadable agent software for GPU owners
- Scheduling: Distributed job scheduling with automatic sharding
- Security: Session-based auth, sandboxed execution, data isolation

## Why Choose Huginn Over Traditional GPU Cloud Platforms

### All-in-One Training Experience
- Other platforms offer raw GPU rental — Huginn provides an integrated AI training pipeline
- 7 free base models included — no need to download or configure separately
- Built-in Model Studio UI with 4-step training wizard
- ChatGPT-style playground to test your model instantly after training
- Automatic cost estimation with upper bound guarantees — no surprise bills

### Universal Device Network
- Unlike datacenter-only platforms, Huginn pools compute from any device worldwide
- Desktop GPUs, laptops, tablets, phones, EVs, and browser tabs all contribute
- Lower entry cost for both customers and compute providers
- Community-powered network with transparent pricing

### Key Differentiators
- Fixed HU pricing (€0.21/HU) vs auction-based or hourly GPU rental
- Persistent training jobs with no session limits or timeouts
- Production-ready model output (.safetensors) ready for local or cloud deployment
- Integrated billing, monitoring, and model management in one platform

## Common AI Training Use Cases on Huginn

### Custom Chatbot Training
- Train a chatbot that speaks your brand voice using your FAQ data
- Fine-tune LLaMA or Mistral with customer service conversations
- Deploy via Huginn Playground or download for local use

### Code Generation AI
- Fine-tune CodeLlama 7B on your codebase for project-specific code completion
- Support for Python, JavaScript, SQL, and more
- Ideal for internal developer tools

### Multilingual AI
- Train models that understand and generate text in 100+ languages
- LLaMA 3.1 and Qwen 2.5 excel at multilingual tasks
- Perfect for global businesses needing localized AI

### Sentiment Analysis & Classification
- Fine-tune small models (Phi-3 Mini) for fast classification
- Customer review analysis, support ticket routing, content moderation
- Low cost: typically 2-5 HU (€0.42-1.05)

### Medical, Legal & Financial AI
- Domain-specific fine-tuning for specialized vocabulary
- Full fine-tune recommended for maximum accuracy
- Data stays private with sandboxed training on Huginn

## AI & Machine Learning Glossary

- **LLM (Large Language Model):** Neural network trained on text to generate human-like responses
- **Fine-Tuning:** Adapting a pre-trained model to a specific task using custom data
- **LoRA:** Low-Rank Adaptation — efficient fine-tuning that trains only small adapter matrices
- **QLoRA:** Quantized LoRA — 4-bit quantization for lower memory usage
- **VRAM:** Video RAM on GPU — determines how large a model you can train
- **Transformer:** Neural network architecture using self-attention, used by all modern LLMs
- **Inference:** Using a trained model to generate predictions/text
- **Epoch:** One complete pass through the training dataset
- **Learning Rate:** Controls how much the model adjusts per training step
- **Batch Size:** Number of examples processed simultaneously
- **Overfitting:** When a model memorizes training data instead of learning general patterns
- **Token:** Smallest unit of text processed by an LLM (roughly 0.75 words)
- **Embedding:** Dense vector representation of text in high-dimensional space
- **RAG:** Retrieval-Augmented Generation — combining search with LLM generation
- **GGUF:** Model format for llama.cpp, optimized for CPU/GPU inference
- **safetensors:** Safe model serialization format used by HuggingFace and Huginn

## Frequently Asked Questions for Search

### How to train an AI model without coding?
Huginn's Model Studio provides a no-code 4-step wizard. Upload your data, select a model, configure settings, and launch training — no programming required.

### What is the cheapest way to fine-tune an LLM?
Use Huginn with QLoRA on T4 tier GPUs. A 7B model with 1,000 examples costs approximately €1-2 (5-8 HU at €0.21/HU).

### Can I train AI on my own data?
Yes. Upload your data in JSONL, CSV, or Chat format to Huginn. Your data stays private and isolated. Maximum file size is 500MB.

### How to make money with GPU sharing?
Install the Huginn Agent on any device — desktop, laptop, tablet, phone, or even an electric vehicle. Share idle compute and earn EUR via Stripe. Desktop GPUs earn up to €0.60/hour.

### What is distributed GPU training?
Distributed GPU training splits AI workloads across multiple GPUs in different locations. Huginn's network pools consumer and datacenter GPUs into a unified compute grid, reducing costs by up to 50%.

### Is AI training bad for the environment?
Traditional datacenter AI training has a significant carbon footprint. Huginn reduces this by reusing idle consumer hardware instead of powering dedicated datacenter GPUs, saving approximately 0.092 kg CO₂ per HU processed.