@BenPielstick@kimmonismus Yep, runs smoothly on RTX 3090, but the model think too much, wrote the whole essay on how many “r” in strawberry.
On some prompts runs out of output tokens and cuts the answer. Seems “smart” though, much smarter than DeepSeek R1 32B, which I’ve tried earlier.
@BenPielstick@kimmonismus Looks like bullshit, the full model weights just 20 GB, a single RTX3090 would be enough, I have one, so I can try installing the model and have a test. Will write later about results.