I pay Moonshot $100💸a month for Kimi K3, and I still ration its tokens. vLLM can’t use an FP8 KV cache on Ampere for its QSA path. My GPUs don’t have native FP8 support, so there are 4 specific spots in the code where everything crashes. K3 read the code, found all 4, and added the missing decoding path. The first version ran 22.8 times slower than BF16. He didn’t care about that at all and didn’t stop. He figured out why on his own, rewrote the tiling, ran a sweep through the configurations, and got the kernel to 1.05x BF16 speed on decode and 1.45x on prefill. The KV cache grew from 104,425 tokens to 178,913 in the same memory. Awesome. I’d periodically walk over to the Mac to see how things were going. In another session that same day, my W4A16 build was spitting out the same garbage string on every run. K3 managed to figure it out and traced it back to a single missing 2-byte scalar in the 180B model. It keeps missing with its own edit tool. I don’t give a damn, honestly. I’ve never said this about a model before. I love this thing. Very much. So @Kimi_Moonshot, my only complaint isn’t with the model at all... The limits on Coding Plan for $100 are way too low. I burn through them and then I just wait. I’d run K3 3x as much if you let me. For now, I’m saving it. I’m reserving it for kernels, inference engines, and those runs that really matter, and it hasn’t let me down once. Thanks, Kimi - my heart is with you ❤️