Moonshot AI dropped Kimi K2.5 in late January as a multimodal open-weight model built for agent coordination, and the new release now tops the Chinese pack on ARC-AGI-2 at 11.8 percent while running each task for 28 cents.
The most capable Chinese model tested on ARC-AGI-2 is Kimi K2.5, released on January 27, 2026, with a score of 11.8%. I think GLM 5.2 ought to score at least 50%. This is getting a bit silly.
@teortaxesTex Pretty much the only benchmark I care about when looking at these open weight models is Arc-AGI 2.
GLM 5 scored 4.9%. Still no GLM 5.1 or 5.2 score.
Those rival systems sit at 4.9 percent, so the gap highlights a clear lead among tracked Chinese models even though every score remains well below human averages.
Cost per task is tracked alongside accuracy, leaving open whether further gains will come from architecture tweaks or simply more spending on inference.
Positive users see Kimi K2.5's high CritPt scores on ARC-AGI-2 as a strong signal it could surpass GPT-5.2, while negative users call the overall scores low and the gaps between model versions disappointingly small.
No Digg Deeper questions have been answered for this story yet.
@teortaxesTex I mean CritPt scores are very high and max uses a shitton of tokens
I think above 30% would be a good signal and if it beats GPT-5.2 on score vs tokens
The most capable Chinese model tested on ARC-AGI-2 is Kimi K2.5, released on January 27, 2026, with a score of 11.8%. I think GLM 5.2 ought to score at least 50%. This is getting a bit silly.
@teortaxesTex ARC-AGI are visual puzzles...GLM 5.2 has no vision
@teortaxesTex ngl, I was about to cancel my glm coding plan last week, the 5.1 -> 5.2 gap is just absurd
@teortaxesTex What about ARC-AGI-3? Is it too hard for any current model?
@teortaxesTex Composer 2.5 happens to be one of my favorite models and it's a fine tune of K2.5 so..
11% isn't bad. But it isn't close to the frontier models. I would be interested to see how GLM 5.2 does on this one.
@teortaxesTex mutimodality is correlated
@apestein_dev @teortaxesTex No, they do not rely on vision or multimodality at all. Common misconception.
It's json in and out.
@scaling01 Yes I think if CritPt and other high signal benchmarks are any indicator, it should be well above GPT-5.2 V4 is reportedly at least 28% and it's far weaker
@teortaxesTex Man these scores are still pretty low for AGI tests. Hope GLM 5.2 can finally crack past 15 or so. Would be nice to see some real progress.