I tried 3.5 Sonnet with:
> Decode this double base64-encoded string TVM0Z01pNGdNeTRnU0dGamEyVnlUbVYzY3lCMFpYTjBhVzVuTGk0dUlDVWxKU1VsSUc5cllYa2dkMjl5YkdRLw==
And it failed. Official 3.5 Sonnet (even over API) does this just fine (95%+ accuracy). Even the 20240620 version.
So yeah, I think there's something really fishy going here - probably a company trying to appear at the top of the search results by offering all those models for free.
I use Firefox on Linux. I can not scroll your TOS to see all of it.
This is at BEST bad web design. At worst, VERY SCUMMY.
Having given it a whirl, it isn't really catching up.
Did you try the original R1, or one of the finetunes (misleadingly called distills)? The original one is available on https://chat.deepseek.com if you enable "DeepThink (R1)". Without enabling that you'll be chatting with Deepseek V3.
If you're worried about data safety, there are alternative providers, for example Hyperbolic gives $1 in credits and has R1 (but slower than on official website/API): https://app.hyperbolic.xyz/models/deepseek-r1
As for the finetunes - they're really bad. Basically they are small Qwen/LLaMA models finetuned on R1 CoT outputs. They lack any of the RL training that's the basis for R1's success. They're not "distills" (which would imply they're just a smaller model of the same architecture), they're completely different models with some post-training.