Private
Conversations stay on the device.
Nothing is sent to the cloud. Chat history lives locally and can be cleared from the sidebar with Wipe history.
You are using a local model. Nobody can see what you do.
Free, local AI in a browser tab
Nothing you type leaves your machine.
Gratis runs open AI models inside a browser tab, on this device. No install and no account. After a one-time model download, it works offline.
No accountOffline after setup
Runs inChromeEdgeFirefoxBrave
The model comes to the machine. The prompts stay on it.
A cloud assistant processes each prompt on a server somewhere else. Gratis downloads the model once, from Hugging Face, then runs it on this device's own memory and graphics.
Prompts, replies and attached images stay in the tab. There is nothing to sign up for and nothing to install.
Sized to fit
A model that fits, suggested first.
Gratis reads the device's memory and recommends a starting model. Download size, memory needed and reply speed show before anything downloads.
Fits your 16 GB computer with room to spare
Offline
One download, then no connection needed.
Each model downloads once. After that, Gratis works without internet, on a train, on a plane or on a network that blocks AI services.
This is a one-time download. After this, everything works offline.
Free
No account, no install, 11 languages.
Open it in Chrome, Edge, Firefox or Brave and start. The interface runs in English, Español, Français, Deutsch, Italiano, Português, Nederlands, Polski, Русский, 日本語 and 中文. A desktop app is also available.
Why models of real size run in a tab
4GiB→16GiB
Browsers cap WebAssembly memory at 4 GiB, which is below what a useful model needs. TPN Labs forked the wllama inference runtime onto 64-bit memory addressing. That raises the working ceiling in the tab to 16 GiB.
Safari does not support 64-bit WebAssembly memory, so Gratis needs Chrome, Edge, Firefox or Brave.
Every model shows its size before it downloads
| Model | Quant | Download | Needs memory | Reply speed | Input |
|---|---|---|---|---|---|
| MiniCPM5 1B | Q4_K_M | 688 MB | 1.2 GB | ~6.4 tok/s | Text |
| Qwen 3.5 0.8B | Q5_K_M | 798 MB | 1.4 GB | ~7.4 tok/s | Text, photos |
| Qwen 3.5 2B | Q4_K_M | 2.0 GB | 2.7 GB | ~3.8 tok/s | Text, photos |
| Qwen 3.5 4B | Q4_K_M | 3.5 GB | 4.2 GB | ~1.9 tok/s | Text, photos |
| Qwen 3.5 9B | Q4_K_M | 6.7 GB | 7.5 GB | ~1.0 tok/s | Text, photos |
As listed by Gratis on a 16 GB computer. Reply speed is Gratis's own estimate for that device and changes from machine to machine. The full list has more models, filters for photos, reasoning, multilingual and long documents, and a field for a custom Hugging Face model.
Chats and images stay on this device. Only the model download comes from Hugging Face.
FAQ
Is Gratis free?
Yes. There is no account, no subscription and no sign-up. Open the page and pick a model.
Does anything leave the computer?
Conversations and images stay on the device. Nothing is sent to the cloud. The only thing that crosses the network is the model download itself, which comes from Hugging Face.
Which browsers work?
Chrome, Edge, Firefox and Brave. Safari does not support 64-bit WebAssembly memory, which Gratis needs to load models of real size.
What Gratis doesn't do YET
- No local API for other appsGratis runs inside the browser sandbox, so it cannot serve other applications. It is for using a model. TPN Bench is for building with one.
- No web researchModels answer from what they already know.
- No SafariSafari lacks the 64-bit WebAssembly memory Gratis depends on.
How much memory does a model need?
Each model lists its download size and the memory it needs before anything downloads. MiniCPM5 1B needs 1.2 GB; Qwen 3.5 9B needs 7.5 GB. Browsers may hide or limit RAM information, so Gratis lets the user say how much RAM the device really has.
Does it work offline?
Yes, after the initial setup. Once a model has downloaded, Gratis works without internet.
Can it read photos and screenshots?
Models marked Photos understand photos and screenshots, for example the Qwen 3.5 models. Others read text only. The list shows which is which.
Can I load my own model?
Yes. The model list has a field for a custom Hugging Face model, in GGUF format.
Is there a desktop app?
Yes. The Install app link on the first screen installs Gratis as an app. It runs the same models, locally.
Who makes Gratis?
TPN Labs, the team behind TPN, Bittensor Subnet 65, and TPN Bench. Several of the models in Gratis are published by TPN Labs on Hugging Face.