Run Ministral 3 3B Reasoning 2512 in your browser

1.8 GB for the q4f16_1 build, in 63 files. It downloads once and stays in your browser cache, so every later visit starts straight away and works offline.

A browser with WebGPU — recent Chrome, Edge, Firefox or Safari — and about 2.8 GB of GPU memory. The tool checks your GPU, memory and free space and tells you whether Ministral 3 3B Reasoning 2512 fits before anything downloads.

No. Ministral 3 3B Reasoning 2512 runs on your own GPU inside the browser tab. The only network traffic is the one-time download of the model files from Hugging Face; your prompts and the answers never leave the device, and the chats are stored in this browser.

Start with q4f16_1 — it is the smallest download and the fastest to run. The f32 builds are larger and use more memory, but they run on GPUs without 16-bit shader support. Builds marked 1k hold a shorter conversation in exchange for less memory.

Its context window is 4k tokens. Longer conversations keep working — the oldest turns are dropped with a notice when the conversation no longer fits.

Ministral 3 3B Reasoning 2512 always thinks before it answers. The reasoning streams into a collapsible block above the answer and is never sent back as part of the conversation.

The weights are Apache 2.0, taken from the model card of mistralai/Ministral-3-3B-Reasoning-2512. Check the licence yourself before using it commercially.