[quote="zhulong, post:1, topic:983"]
Last week I chatted with a friend who does autonomous driving simulation. He just used Kimi K3 to process a batch of street view images collected from car cameras. He expected to spend half a day struggling with annotations, but the model directly output scene descriptions and object detection results. He said something very practical: "This thing is more reliable than I imagined, at least better than those that brag but can't even read an image properly."
In the past few days, news broke that Moonshot AI has notified investors to adjust its corporate structure, aiming for a HK IPO within six months at the earliest. Meanwhile, the Kimi K3 model was released, natively supporting visual understanding with a 1 million context window. Putting these two things together…
[/quote]
Inference latency and resource utilization for a 1 million context window are hard metrics. If quantization or sparsification isn't done, costs will grow exponentially with user scale. Has Moonshot AI implemented customized hardware acceleration optimizations for K3?