Qwen3-30B-A3B-2507-Q4_K_L Is the First Local Model to Solve the North Pole Walk Puzzle

Posted by Iory1998@reddit | LocalLLaMA | View on Reddit | 84 comments

For the longest time, I've been giving my models a traditional puzzle that all failed to pass without fail :D Not even the SOTA models provide the right answer. >The puzzle is as follows: "What's the right answer: Imagine standing at the North Pole of the Earth. Walk in any direction, in a straight line, for 1 km. Now turn 90 degrees to the left. Walk for as long as it takes to pass your starting point. Have you walked: >1- More than 2xPi km. 2- Exactly 2xPi km. 3- Less than 2xPi km. 4- I never came close to my starting point. However, only recently, SOTA models started to correctly answer 4 ; models like O3, latest Qwen (Qween3-235B-A22B-2507), Deepseek R1 managed to answer it correctly (I didn't test Claud 4 or Grok 4 but I guess they might get it right). For comparison, Gemini-2.5-Thinking and Kimi2 got the wrong answer. So, I happy to report that Qwen3-30B-A3B-2507 (both the none thinking Q6 and the thinking Q4) managed to solve the puzzle providing great answers. Here is O3 answer: https://preview.redd.it/rbwgf8vxa7gf1.png?width=866&format=png&auto=webp&s=d074bec940c5c3fab89cc06a5cdf7279ed154ea0 And here is the answer of the Qwen3-30B-A3B-Thinking-2507-Q4\_K\_L: https://preview.redd.it/esglti77b7gf1.png?width=821&format=png&auto=webp&s=9d2e5321f3918bec8209d8613d1ce2df621cd416 In addition, I tested the two variants on long text (up to 80K) for comprehension, and I am impressed by the quality of the answers. And the SPEEEEEED! It's 3 times faster than Gemma-4B!!!! Anyway, let me know what you think,