Free AI for Open Source | Free.ai

Yes. Every model we run on our GPU servers is open-source under permissive licenses -- Apache 2.0, MIT, or BSD. We deliberately exclude models with restrictive licenses like Meta's Llama license. You can inspect, use, and modify all models freely.

Meta's Llama license includes restrictions on commercial use for large companies and requires specific attribution. We only use models with truly permissive licenses (Apache 2.0, MIT, BSD) that allow unrestricted commercial and personal use by anyone.

Every model in our table links to its source on HuggingFace or GitHub. You can download the weights directly from these repositories and run them on your own hardware.

Qwen2.5-72B is our most capable chat model for complex reasoning and analysis. Qwen2.5-7B and Mistral-7B are faster options for everyday tasks. Phi-3 is excellent for efficient reasoning and math.

We use five text-to-speech models: Kokoro (Apache 2.0), Piper (MIT), MeloTTS (MIT), Chatterbox (MIT), and OuteTTS (Apache 2.0). Each has different strengths in voice quality, speed, and language support.

FLUX.1-schnell for fast generation (Apache 2.0), Stable Diffusion XL for high-quality images (OpenRAIL++), and Kandinsky 2.2 for artistic styles (Apache 2.0). For video, we use CogVideoX-2B and AnimateDiff.

If you have developed or know of an open-source model under a permissive license that you think we should host, contact us. We evaluate models based on quality, license compatibility, and user demand.

Active open-source contributors and maintainers can apply for free Pro access to Free.ai. If you maintain a popular open-source project, contact us with a link to your work and we will upgrade your account.

We evaluate every model before deploying it based on benchmark performance, real-world output quality, inference speed, and license compliance. Only models that meet our quality bar are added to the platform.

We use MadLAD-400 3B from Google, which supports 450+ languages under an Apache 2.0 license. It handles even low-resource languages that many commercial translation services do not support.

Real-ESRGAN (BSD) for image upscaling, BRIA RMBG 2.0 (Apache 2.0) for background removal, SAM (Apache 2.0) for image segmentation, and DeOldify (MIT) for photo colorization.

Not all of them. We provide access to 346+ models from various providers, including proprietary models like GPT-4o, Claude, and Gemini. These are proxied through their respective providers and have their own licensing terms separate from our self-hosted open-source models.