K2-Vendor-Verifier, a dead-simple harness that fires the same 2 000 tool-calling prompts at every vendor who claims to sell “Kimi K2” endpoints. The delta is brutal: the best node returns 99.3 % token-perfect answers, the worst barely hits 64 %—same brand, same price, totally different model under the hood.

MegEngine team speedup depthwise Conv 10 times faster, which makes increasing kernel size almost cost-free while gaining better neural network accuracy.

MegEngine (A Deep Learning Framework) implements DTR. Now you can train 3x larger model by tradeoff a little bit speed for lots of GPU memory.