Unify heterogeneous GPUs (2GB–32GB+) into a single compute pool. Train AI models at lower cost with measurable sustainability impact.
One network to train your AI — cheaper, private, and built for the physical world.
🔒
You choose the protection level on every job: Shield+ keeps your data encrypted and split so no machine ever sees the whole — or Vault runs on sealed hardware no one else can read.
More than fine-tuning. Hugin Learning trains robot, drone and automation control from scratch — our own breakthrough, for the problems others can’t train.
Every result is independently re-checked, and a slow or dropped machine never stalls your job. You get a model you can trust — completely hands-off.
What you can build
Bring your data, train production-ready models across the network, then download and use them anywhere.
Fine-tune open models (LLaMA, Mistral, Phi, Gemma, Qwen) on your own data with LoRA & QLoRA.
Chatbots · Assistants · Code · Multilingual
Train image classification & object-detection models on your own images, then download the trained model.
Quality control · Drones · Cameras · Inspection
Build perception models and prepare data — augmentation, validation and benchmarking for robotics teams.
Perception · Dataset prep · Model evaluation
Train control policies for robots and agents — the system discovers behaviour from scratch by trial-and-error, no labelled data needed.
Robots · Control · Agents · Automation
Six-step pipeline from job definition to verified delivery with transparent billing.
Data classification, target metric, and pool selection.
HU & GPU-seconds estimate with upper bound and SLA.
Capability-aware distribution, node selection, redundant execution.
Sandboxed micro-shard execution with telemetry, health checks, auto-retry.
Quality threshold, spot-check, and anomaly detection validation.
Verified job billing, automatic GPU owner payout.
The journey
You upload, we take care of everything, and your finished model is ready to download.
Upload your data
Secure & private
We prepare it
Fully automatic
Trained across our network
Powered at scale
Turned into your model
Brought together
Download the result
Ready to use
Our breakthrough
Our own innovation: control intelligence that learns from scratch — by trial-and-error, with no labelled data — trained and verified across the network. Built for robots, drones and automation.
Discovers the right behaviour by trial-and-error — no demonstrations or labelled data required.
Balancing, positioning, navigation and your own custom tasks — the brain that runs robots, drones and automated lines.
Trained across the network and independently re-checked, so the result can be trusted.
⬡ Engineered by Hugin
Most GPU networks ship your raw data to strangers' machines. Hugin doesn't. Every job runs under Hugin Shield+ — a multi-layer protection stack we built — or, for the most sensitive work, inside a hardware-isolated Vault.
🛡️ Hugin Shield+
Included on every job
Protection is built in on every job — at no extra cost. Your data stays private, safe, and completely yours.
🔒 Hugin Vault
Maximum confidentiality
Maximum confidentiality. Your data stays invisible — even to the computer that trains it.
You choose the protection level per job — Shield+ is included at the standard rate; Vault adds hardware isolation for the most sensitive data.
HU (Hugin Unit) is your simple billing unit. 1 HU ≈ 3,600 normalized GPU-seconds. Pre-estimate and upper bound before every job.
0.21EUR / HU
Transparent Pricing for Training
Share Your Compute
Turn idle devices into passive value. Hugin network harnesses compute power from any connected device globally.
| Device Class | ~HU/Hour | Examples |
|---|---|---|
| Smartphones (Idle) | 0.05 - 0.15 | iPhone 15 Pro, Pixel 8 |
| Electric Vehicles (EV) | 0.2 - 0.4 | Tesla MCU, Polestar 2 |
| 4-6GB Consumer | 0.4 - 0.6 | GTX 1650, RTX 3050 |
| 8-12GB Consumer | 0.6 - 0.9 | RTX 3060, RTX 4070 |
| 16-24GB Workstation | 1.6 - 2.8 | RTX 3090, RTX 4090 |
| 24-48GB+ Datacenter | 2.8 - 4.5 | A100, H100 |
Everything you need to know about training AI models on Huginn.
Project Huginn is a distributed GPU sharing platform that lets you train and fine-tune AI models at up to 50% lower cost than traditional cloud providers. It pools heterogeneous GPUs (from RTX 3080 to H100) into a unified compute grid. Built by SMware ApS, a Startup Denmark approved company.
Huginn uses Hugin Units (HU) for billing. 1 HU = €0.21. A typical LoRA fine-tuning of a 7B model with 1,000 training examples costs approximately 5-8 HU (€1-2). Costs are estimated before you start, with an upper bound guarantee — no surprise charges.
Huginn provides 7 base models for free: LLaMA 3.1 (8B & 70B), Mistral 7B, Phi-3 Mini 3.8B, CodeLlama 7B, Gemma 2 9B, and Qwen 2.5 7B. You select a base model and upload your own training data — Huginn handles the rest.
Three methods: LoRA (Low-Rank Adaptation) for fast, efficient training with ~12GB VRAM; QLoRA (4-bit quantized) for even lower memory usage with ~8GB VRAM; and Full Fine-Tune for highest quality with ~40GB+ VRAM. LoRA with rank 16 is recommended for most use cases.
JSONL (Alpaca format with instruction/input/output fields), CSV (with instruction, input, output columns), and Chat Format JSONL (with messages array). Maximum upload size is 500MB. We recommend at least 1,000 examples for good results.
Unlike traditional GPU rental platforms that only offer raw compute, Huginn provides an all-in-one AI training experience. You get a built-in Model Studio with 7 free base models, an integrated fine-tuning pipeline, automatic cost estimation with upper bounds, and a ChatGPT-style playground to test your model instantly — no setup, no configuration, just upload your data and start training.
Yes! Huginn lets you earn from virtually any device. Share your desktop GPU, laptop, tablet, smartphone, or even your electric vehicle's compute power while you drive. Connect via the Huginn Agent app or simply open a browser tab. Every connected device contributes to the network and earns you HU. Desktop GPUs earn the most (up to €0.60/hour for high-end cards), but even a smartphone or browser tab generates passive income 24/7.
It takes under 5 minutes: 1) Sign in with Google and choose GPU Owner role, 2) Go to GPU Agent page in your dashboard, 3) Download the agent file and generate a token, 4) Run the command in your terminal. The agent auto-detects your GPU and starts processing jobs. You can also use the browser-based Web Engine with zero downloads. Read our full guide at projecthuginn.com/how-to-share for step-by-step instructions.
Yes. All training jobs run in sandboxed environments with complete data isolation. Each user's models and datasets are private. We use session-based authentication and encrypted connections for all data transfers.
Yes. After training completes, you can download your fine-tuned model in .safetensors format. You can run it locally with Ollama, llama.cpp, vLLM, or HuggingFace Transformers. You can also keep it on Huginn and test it via the built-in playground.
Five tiers: T1 (H100/A100 80GB+, 4.50 HU/hr), T2 (A100 40GB/RTX 4090, 2.80 HU/hr), T3 (RTX 4080/3090, 1.60 HU/hr), T4 (RTX 3080/2080 Ti, 0.90 HU/hr), and T5 (entry-level, 0.40 HU/hr). Higher tiers are faster but cost more per hour.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。