浏览器方块游戏：Kimi 与 Astra 对比

原帖说明

SAME GAME, TWO MODELS. I'M MOVING MY BUILDS FROM GPT-6 ASTRA TO KIMI K3.1 PRO, AND HERE'S WHY. Kimi K3.1 Pro dropped today. the spec sheet reads like it was written for people who build games: → full-project coding in Unreal, Unity and Godot → 1M tokens of context, a whole codebase in one conversation → 2.8T parameters, only 104B active per token → a Highspeed variant, 2 to 3 times faster by Kimi's numbers → open weights promised by November 4 with Astra I paid for the best answer and rented it. with Kimi I'll own a copy of the model itself once the weights land the first build I'm moving over is a blocky game world: village, blacksmith, mines, nether, a portal at the end. what decides it for me: → fewer retries on the same bug → a project that still makes sense at file fifty → whether the 1M context actually holds the whole project or just looks good on the launch page → how often it changes things I didn't ask it to touch → code I can read without asking the model to explain it it might be start of the end for OpenAI

— @pilvex