The Lab
Reading about models is fine. Touching them is better. Everything here is a real Meta model — running in your browser or precomputed on the Raspberry Pi that hosts this site. No cloud calls, no API keys, no tricks.
Segment Anything
Click any object and the actual SAM mask decoder — 8-bit quantized, running in WebAssembly — cuts it out in real time. The same architecture as Meta's own demo.
✂️Cut anything out of your photo
Drop in your own picture and click: Meta's EfficientSAM encoder and decoder run in your browser (11 MB, once). Nothing is uploaded — save the sticker.
🌈How DINOv2 sees
The paper's famous PCA rainbows on our own photos, plus a live patch-matching toy: hover a retriever's ear, light up a dalmatian's.
📷Your photo, through DINOv2
Drop in any photo — the real ViT-S/14 runs in your browser (17 MB over the wire, once) and paints the PCA rainbow live. Nothing leaves your device.
🔤How a model reads
Type anything and watch it become tokens — OPT-175B's and Llama 3's real tokenizers side by side, in your browser. See what Hindi, code and emoji cost, why Llama 3 grew to 128K pieces, and why BLT wants to abolish tokens.
🌍200 languages
Our sentences through NLLB-200's eyes — every FLORES-200 language, every script, instant switching. Then test yourself: guess the language.
🎵MusicGen jukebox
Ten prompts, ten compositions — sitar sunrises to chiptune arcades — generated by musicgen-small on a Raspberry Pi, with a live spectrogram.
How is this possible on a static site?
Big models split naturally into a heavy understanding half and a light interactive half. We run the heavy half once, offline, on a Raspberry Pi 5 (SAM's image encoder, DINOv2's features, NLLB's translations, MusicGen's synthesis) — and ship the light half to you (SAM's prompt decoder — 4.7 MB on disk, 3.2 MB over the wire — in WebAssembly, cosine similarity in plain JavaScript). It's the same trick Meta's original Segment Anything demo used in 2023, and it's why a hobby server in Mumbai can hand you a frontier lab's models with zero cloud compute.
Three demos go further and run the whole model in your tab, because the model is small enough: EfficientSAM's encoder and decoder for your own photos, DINOv2's ViT-S/14 for your own photos, and both tokenizers, which are pure arithmetic over a vocabulary file. Once those files have downloaded, your pictures and your text never leave the device, and the work carries on with the network switched off. Reloading the page does need the network again — none of this is installed, it is just held in the browser cache for the session.
Weights belong to their creators; open licenses noted on each page. This site is unofficial and not affiliated with Meta.