Free AI Hosting
Host AI models for free. GPU access, API hosting, and cloud deployment.
Cloud Hosted
Use Free.ai infrastructure. Zero setup, zero maintenance. All models are pre-loaded and ready to use via API or web UI.
Available NowSelf-Hosted (DIY)
Run the same open-weight models on your own hardware, using the weights their authors publish and standard serving tools. We do not publish a Docker image.
Self-ServiceManaged Private
Dedicated GPU servers managed by us, deployed in your preferred cloud region. Full data isolation and custom SLA.
EnterpriseSelf-Hosted Deployment
Our self-hosted models are open-weight models whose authors publish the weights, most of them under permissive licenses (check each model card). We do not publish a Docker image or our own inference server; to run a model yourself, use a standard open-source server such as vLLM:
# Example with vLLM (third-party, Apache 2.0) and a Qwen model from Hugging Face
pip install vllm
vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507 --port 8000Minimum Requirements
- NVIDIA GPU with 24GB+ VRAM (RTX 4090, A5000, A100)
- CUDA 12.0+ and a recent NVIDIA driver
- 16GB+ system RAM, 100GB+ storage per model
- For 72B parameter models: 80GB VRAM (A100) or multi-GPU setup
Why Self-Host?
- Data privacy — Your data never leaves your servers
- No rate limits — Unlimited inference on your hardware
- Compliance — Meet data residency requirements
- Customization — Fine-tune models on your data
- Cost control — Fixed hardware costs, no per-token fees
- Air-gapped — Runs fully offline
FAQ
Cloud Hosted (use our infrastructure, zero setup) is the standard option. You can also run the open-weight models yourself on your own GPUs with standard tools, without our help or our software. A managed private deployment (dedicated GPU servers run by us) is scoped and quoted per account.
It depends on the model and the serving tool. As a rough guide, an NVIDIA GPU with 24GB+ VRAM (RTX 4090, A5000, A100), CUDA 12.0+, 16GB+ system RAM and 100GB+ storage per model; 70B-class models need 80GB VRAM or several GPUs. Check the model card and your serving tool's documentation.
If you run open-weight models yourself, they can run offline once you have downloaded the weights and your serving tool. That setup is yours: we do not ship software for it.
We do not publish a Docker image or an installer. Download the model weights from the model author (usually on Hugging Face) and serve them with a standard open-source tool such as vLLM. If you want us to deploy and run it for you, email hello@free.ai and we scope it per account.
All self-hosted models use permissive open-source licenses -- Apache 2.0, MIT, or BSD. You can use them commercially without restrictions. We deliberately exclude models with restrictive licenses like Meta's Llama license.
Managed private hosting gives you dedicated GPU servers in your preferred cloud region, fully managed by our team. We handle setup, patching, model updates, and monitoring. You get full data isolation with an enterprise SLA.
If the model's license allows it, you can fine-tune the open weights yourself with standard frameworks such as Hugging Face Transformers. We do not provide fine-tuning tools or images.
Contact our sales team to discuss a trial period. We typically offer a short evaluation period for enterprise prospects to test managed private hosting before committing to a long-term plan.
Cloud hosting uses the standard token-based pricing. Self-hosted is free -- you only pay for your own hardware and electricity. Managed private hosting is priced based on GPU allocation, region, and SLA level.
Yes. You can run some open-weight models yourself and use the Free.ai cloud for everything else. Routing requests between the two is up to you, and your own server's API will follow whatever tool you run.
We do not provide software or support for do-it-yourself deployments. A managed private deployment, scoped per account, includes technical support by email, monitoring and a dedicated account manager.
Cloud hosted is best for teams that want zero maintenance. Self-hosted is ideal for data privacy, compliance, or unlimited usage on your own hardware. Managed private is the best of both worlds -- full data isolation with no operational burden.