KIMI K2.6 SOON !!

Lemme join the glaze because 2.5 is the one I enjoyed the most while running a personal test. My personal test is: I send 4 puzzles to the AI and tell them to estimate how much it would take for an average person to solve them after they themselves solve it to gauge the difficulty, but to not tell me the result or give any hints. 3 of the puzzles are pretty straightforward, while one (the third one) is really difficult, puzzle 1, 2 and 4 are all related to each other, so if you get one, the others become easier even if they difficulty is increased. MInmax was the worst, it was clear while reading the reasoning that it didn't care about the rules of the puzzles 1, 2 and 4, so it got them wrong by disobeying the limitations, and also gave up on the third one, then it told me the wrong results it found despite the direct instruction not to. Deepseek, GLM 5.1 and Gemma 4 solved 1 and 2, skipped 3 after trying for a while and went to 4, they did obey the instructions. Kimi initially also skipped 3, but then came back to it after estimating the times for 1, 2 and 4 and kept trying until it got it right, it also impressed me while reading it's reasoning that it made the following comment before skipping number 3 "Given the difficulty I'm having solving it (and I have computational advantages), I'd estimate this takes the average person hours or days, or they simply won't solve it.", I really liked that it acknowledged that humans can't crunch info as fast as LLMs, I also found it charming that after skipping number 3 and solving number 4, it said. "Wait, let me reconsider puzzle 3. Maybe there's a clever insight that makes it easier?", and then went back to it until it actually managed to solve the puzzle, concluding "Given how long it took me (with systematic analysis), this is indeed very difficult. An average person without a systematic approach would likely struggle for a very long time. I'd estimate 1-3 hours for a very determined person, or they might not solve it at all without hints."