TEST KIMI K2.6 1T - A32B: The BEST Open AI?

Hinzugefügt: 2026-05-05

[00:00:00]Hello community, so great that you are back. We have a moon shot and you Kimi K 2.6 model. Isn't this beautiful? So, let's test it out. Here you can go Kimi.com blog Kimi K 2.6 and you have all the technical details, but you know we are interested simply in the performance. Now, there's one benchmark that show us, hey, the jump from 2.5 to 2.6 is really significant and it is real right behind our trio here that has here on the artificial analysis index here all 57 points. This is Claude Opus 4.7 Max, Gemini 3.1 Pro preview and GPT 5.4 X high.

[00:00:44]This is the leading group, but then Kimi K 2.6 and we just ask, is this possible?

[00:00:51]Especially if you look here at the price point, Kimi seems to be really really an interesting model. So, let's examine about the performance of Kimi and we do a causal reasoning test logic long reasoning traces. Let's go. And we are live here at arena.ei and you see on the left side I have Kimi K 2.6. On the right side we have the old Kimi K 2.5 sinking and you see both models started here and you see here the sinking trace or the reasoning trace of both models.

[00:01:24]And you immediately see there's a difference now, huh? At least here in the reasoning performance, Kimi 2.6 is completely different. Look, it is analyzing now here in a different way how the exercise is structured.

[00:01:43]It understands here all the limitations.

[00:01:46]If you look at 2.5, you see it is more or less, hey, can I do both? Let's try this out. Let's go. But you see Kimi 2.6 is here much more strategic. Another sort alternative, what can I think of?

[00:01:59]Given this complexity, the best strategy is to avoid. 2.6 is much more strategically oriented. Look, it looks at the code cards. It analyzes, okay, that's all the instruction. Yellow is never described, so it finds out exactly that there are objects that have not a mathematical function behind it. You know, this is my classical test. This is my elevator test. We have we are on the floor zero. We want to go to the floor 50. The building has only 50 floors and there are some button presses you can press in this elevator that have specific mathematical functions behind each button press.

[00:02:39]We have a time inversion. We have a location inversion and other dependencies. We have interwoven dependencies. We have an energy optimization. We have a token optimization and more. So, this is really here as you can see, those are the reasoning traces of both Kimis.

[00:02:57]And just look at the difference. Yeah, 2.6 is much more strategic. It is much more formulating here its reasoning trace. It's an open reasoning trace. I really can see what's happening and this is real nice.

[00:03:11]Kimi 2.5 sinking seems to be much further advanced now already trying here to find the solution, while 2.6 is still here more or less in the strategic planning phase.

[00:03:24]So, therefore I think 2.6 here will take a little bit more time. Look, start at zero. 2.6 starts now at floor zero. Look at 2.5 sinking. We are already here at floor 31, 35.

[00:03:40]So, there is a massive difference here in the detail in the precision here of the planning before we actually start here anything. So, 2.6 is definitely a different beast. Okay.

[00:03:55]They also have in 2.6 quite long reasoning traces. Look at this. Wow, it is a trial and error. That's really trying out. I only have 50 steps, so it can really try out all 50 steps and find the right solution. Okay.

[00:04:11]But I think 2.5 will have a solution real fast. 2.6 is still here more or less. Okay, let's find a route. Now it is trying to go. Now let's go to 50.

[00:04:24]I think 2.5 is already done.

[00:04:26]So, absolutely interesting here.

[00:04:29]2.6 much more detailed, much more nitty-gritty, much more here absolutely precise what is here exactly the function of each button press, what is the strategy, how do I plan this, what about here the code cards here. Really interesting. Oh, it has the emergency exit. 2.6 found a very nice way to go.

[00:04:49]Okay.

[00:04:52]And I think 2.5 already has a solution.

[00:04:55]Yes, the first solution by 2.5 sinking.

[00:04:58]Unbelievable.

[00:04:59]Nine button presses plus an emergency exit. This is a real good solution by 2.5 sinking.

[00:05:07]Now, we don't know if it is a correct solution, a valid solution, but nine plus an emergency exit is from the amount, from the length of the sequence here, a real good result here 2.5 sinking.

[00:05:21]You see, 2.6 is now building its first sequence. So, there are some massive difference here in the internal reasoning, in the internal logic how it approaches here this.

[00:05:34]2.5 is done. We have nine button presses plus the emergency exit button.

[00:05:39]Solved for 3 minutes and 43 seconds.

[00:05:42]Beautiful. And Kimi 2.6 is just warming up. Sequence attempt number two. Okay.

[00:05:49]So, you see we are really starting here to go from zero to floor 15.

[00:05:54]It has the same subproblem sequencing here. It also tries to go to floor 15. Now we are at floor 36.

[00:06:03]Twice what are the strategies? How can we proceed from here? But it looks much more interesting. 4 minutes 30 seconds later.

[00:06:12]I had a look at this, but you just see here now 2.6 is really fighting. It's really fighting. Look at floor 35.

[00:06:21]It is fighting. 2.6 here has not come up yet with a solution.

[00:06:26]2.5 already has a solution. Nine button presses plus the emergency exit. An excellent solution. 2.6 is still in the try out and error. Look, it now it's trying. Now it's a little bit Yeah, we lost a little bit strategy. How to save?

[00:06:47]Something went wrong. Please try again.

[00:06:49]We have a crash. Kimi 2.6 is gone.

[00:06:53]Let's restart from the very beginning.

[00:06:55]No problem. Everybody gets a second chance. Let me restate the problem clearly. So, I hope there is something in the cache already here for 2.6, but otherwise you see it starts again to understand, trying to understand what exactly is it and 6 minutes and 43 seconds later.

[00:07:14]I hope you don't mind if I just skip ahead because it's just happening more or less the same what we just examined here together.

[00:07:21]It is now trying to come up here with its first solution. Nine button presses is worse. Do we have an eight button press solution?

[00:07:32]It's trying alternative.

[00:07:34]We are still at floor 15. Okay.

[00:07:37]Eight press path. So, 2.6 found a better solution. An eight press path plus exit.

[00:07:44]2.5 has nine presses plus exit.

[00:07:48]But 2.6 is still checking. Look, it's not saying, okay, I have it.

[00:07:53]Checking, checking. What has happened if I do this? What happens if it does? What about press number one? Could we reach floor 15 in two presses? What about two presses to floor 11? It is still not yet satisfied.

[00:08:08]But it seems there's no better solution.

[00:08:10]We do have an eight press solution plus the emergency exit. But it took us here, I would say in total I'm sitting here close to 20 minutes now.

[00:08:20]And we have not No, no. Almost. We almost have the first answer by Kimi K 2.6.

[00:08:28]At about 20 minutes time.

[00:08:31]But you see now, wait, we are still calculating. We are still trying to find a better solution.

[00:08:37]The system is not yet convinced that it is able to answer this.

[00:08:42]We need to land on green. It is verifying again. My goodness. I don't know how many steps at 15 I have seen now. 20, 30 times the same thing. It is really trying, testing, trying, testing.

[00:08:56]My goodness, it does not trust in its own reasoning capabilities.

[00:09:01]Double check, double check.

[00:09:04]Again, look, we go press 5 6 7 8 9.

[00:09:08]Again, we check again.

[00:09:11]It's more or less more than half of the time it is just trying to validate its reasoning trace internally.

[00:09:18]We have not yet gotten an answer by Kimi 2.6. Nothing. Not one bit of an answer.

[00:09:27]Look, it is now checking what numbers are prime.

[00:09:32]Final answer. Should nicely format it.

[00:09:34]We do have an answer.

[00:09:37]I will present this clearly. Yes, do it.

[00:09:39]Give me the answer, please.

[00:09:42]First solution by Kimi K 2.6. Yes, we have it. We have a eight button press plus emergency exit. So, it was worth waiting the time. Close for me now here in real life, close to 20 minutes for the first real answer by Kimik 2.6.

[00:10:03]It looks really nice. I do this now for a year. I can tell you this is absolute correct what it is analyzing here. I think this is here a beautiful solution here. Eight button press plus exit here.

[00:10:16]It is not the minimum. This is not correct. But 2.6 after 9 minutes and 26 seconds Yes, absolutely eight presses and emergency exit is better than 2.5 thinking.

[00:10:29]But I think 2.5 took about 3 minutes.

[00:10:33]So yeah, if you have to pay for the reasoning tokens yeah, you think about it. But yeah.

[00:10:40]2.5 has not the best sequence at the start with DDD. You see the ABAB at 2.6 is much better.

[00:10:49]So let's see if the result can be validated by those models.

[00:10:55]Okay.

[00:10:57]So No, you and I know that 2.6 has the right result. So the first validation run is just to see if the system itself agrees that this is a valid result. And I know you might be laughing about this point, but I can tell you I have seen systems that had a right correct result, but failed in the validation because it just hung up somewhere in a local minimum and it was not able to recover.

[00:11:22]It was not able to provide a final answer if the result can be validated.

[00:11:28]Now we do it.

[00:11:29]Kimik 2.6 real nice, real precise. Look at this.

[00:11:34]Really going here for each step.

[00:11:37]It is if you compare it to the 2.5 here on the right hand side 2.6 is much more precise. But again, look, wait. No, let's check again. No, wait here. Let me check all constraint.

[00:11:50]My goodness, this has an internal checking craziness that is almost unbelievable.

[00:11:57]This system 2.6 does not trust itself at all.

[00:12:02]I think three, four time every single item is checked and rechecked and checked again.

[00:12:08]This is not a good sign if the developer decided that this is a necessary step before we get an answer.

[00:12:17]But it is absolutely precise. On the positive side, 2.6 if you do not have to pay for the reasoning traces here is absolutely precise.

[00:12:26]Or you have it locally and you don't care but you use a little bit more energy. Nice. So up until now I would say absolutely 2.6 is worth that you switch from 2.5 to 2.6. It has a better performance. It takes much longer. The reasoning traces are much more detailed.

[00:12:44]The reasoning traces are much more checked and rechecked and double checked again. But up until now 2.6 has a better performance compared to the 2.5 thinking. And remember we are still in the reasoning traces. We still see the thinking process going on on both models. We have not yet received a valid answer on the validation.

[00:13:06]We are just thinking here and this is nice if we have open source model. We have access here to the reasoning traces.

[00:13:13]And you know, every big company is afraid that from those reasoning traces other startup companies could then distill here some little AI that are almost as intelligent. So you know the proprietary companies here like OpenAI, if you have those tests, they have restriction here on the transparency of the reasoning traces because simply they do not want it any other company gets access to the reasoning. But we have here access to the reasoning and I have to tell you 2.6 solid, yes, absolutely all good. I think the validation is straightforward. I'll present this clearly.

[00:13:52]And now we have a 2.6 the first validation, verification. Now let's have a look. What is the result by 2.6?

[00:14:00]This is now interesting.

[00:14:03]Okay, this looks good. This looks good.

[00:14:06]Button press seven is correct. Yes, button press eight is correct. Yeah, nine emergency exit. Beautiful.

[00:14:13]Yes, we have a validation. Eight plus emergency exit is correct. Look.

[00:14:18]Green check marks all over the place.

[00:14:20]Yes, zero traps. Beautiful.

[00:14:23]Yeah, 2.6 could validate its very good result, an excellent result by 2.6.

[00:14:30]Now it was faster.

[00:14:33]No alternative path could be shorter.

[00:14:35]This is incorrect. There is a better path. But 2.6 is not able yet to do it.

[00:14:41]Validation 2.6 successful.

[00:14:44]Celebration. Beautiful.

[00:14:47]The run is fully valid. It respects every rule, meets every resource, every code requirement. Beautiful.

[00:14:52]The 2.5 is still fighting with its own solution. You see the DDD at the beginning is causing a problem to 2.5 thinking. And maybe it found out that this is not a valid solution.

[00:15:06]So 2.5 came up with something. Yeah, we have a contradiction.

[00:15:11]Okay. So you see the 2.5 thinking. Yeah, it is time to switch to 2.6.

[00:15:18]It is as you see here in this causal reasoning test of a long reasoning trace.

[00:15:27]It better. It has a better result.

[00:15:30]And you see the reasoning process is different. So they really changed quite a lot of jumping from 2.5 to 2.6. We still have not Oh, I need to present this clearly. All things are satisfied. So we have a new solution. 2.5 found a correct solution. Let's see if this is really correct.

[00:15:53]But what about the length?

[00:15:56]I just see that there is a new solution I have not seen yet. So the final sequence is No. Oh, there's an F. No.

[00:16:04]No, it is uh not a shorter solution it found. A solution that is not as good.

[00:16:09]10 button presses plus emergency exit.

[00:16:12]So unfortunately the eight button presses and emergency exit and nine button presses and emergency exit were not able to be validated.

[00:16:22]Constraints, we have a check mark everywhere. Beautiful. Okay, this looks good.

[00:16:30]Yes, now we landed here on the problematic floors. Okay, emergency exit condition is green.

[00:16:36]F is necessary and legal. Well, this is not really what we are looking for.

[00:16:42]And then we have Do we really have 10 plus exit? Yeah.

[00:16:47]So the validation of its original failed, but we came up with a new solution and this is now 10 button presses plus exit.

[00:16:55]So 2.6 so much better.

[00:16:59]Eight button presses plus exit.

[00:17:01]Beautiful. So switch to Kimik 2.6. It is not the best yet, you know. The Gemini 3.1 Pro had the best. So really Kimik is behind the leading triplet here. But eight plus emergency exit is a beautiful solution here for this artificial intelligence system.

[00:17:23]Forget about 2.5 thinking and non-thinking. It is not up to the task as you can see here in this video.

[00:17:32]This one is optimal is not correct by 2.6, but up to the ability of 2.6 it is here the best it could do.

[00:17:40]Now we copy this sequence and say, let's validate this new solution by Kimik 2.5.

[00:17:46]Let's run both models here only on the solution that 2.5 found and I want to see if 2.5 finally came up here with a valid solution. I want to end the test that 2.6 has a validated solution and 2.5 thinking has a validated solution. And yeah, 2.6 thinking you see it is thinking. So I think we can forget about thinking and non-thinking labels at models. All the brand new models are reasoning models if not for very specific cases that they have to run on your phone or other edge devices or on your Apple Watch, whatever.

[00:18:23]2.6 really structured, precise, clear.

[00:18:27]Step eight, B, AB, check, check, check.

[00:18:30]Step nine, C, start, normal, check, check, check. But wait.

[00:18:35]Wait, there was something. No, check.

[00:18:37]Step 10, B. Okay.

[00:18:39]Collect the two cards. Did I miss anything? Let me double check. Anything 2.6 is double checking everything.

[00:18:46]2.5 is Yeah, just go with it. Okay.

[00:18:51]No, step 10.

[00:18:53]Any issues with the deep preset?

[00:18:57]Okay, I see. Now this is Yeah, I forgot to put in the sequence plus emergency exit. Okay, I made a mistake in the valid solution. I would have to end the sequence B plus emergency exit. I forgot the emergency exit. So the effect is now it will land on floor 29 and then take the shortcut to 50. Okay, apart from this detail, let's have a look.

[00:19:22]Yeah.

[00:19:24]It is legal, but yeah, I missed the exit here. Okay, this was my mistake.

[00:19:30]Yeah, but you see otherwise everything is okay.

[00:19:33]The run stops at 29. Therefore it is not a valid solution. I know because I forgot to put in manually here the sequence the emergency exit.

[00:19:42]But it says, yeah, otherwise it is valid.

[00:19:45]And 2.5 tells me, well, it's a partial valid run. It stops at the destination and appended. Yes.

[00:19:53]Okay, we have two validated solution.

[00:19:57]Beautiful.

[00:19:58]And I would like to improve this and I say, "Which strategy would improve your results significantly?" Because we want to reduce the length of the sequence, no? So, try to find a strategy and then explain it.

[00:20:11]So, now we will see if Kimi K2.6 after 5 minutes, you don't have to watch this, comes up with a seven-step solution plus an emergency exit. Can it do the thinking here of a Gemini 3.1 pro or not?

[00:20:27]And what is happening to our K2.5?

[00:20:31]Okay, 2.6 tells me No, it stays with its solution. Okay.

[00:20:38]No, Kimi could not optimize itself, but 2.5 could optimize itself.

[00:20:46]Yes, now I also have here eight button presses plus exit. Both model found eight button presses plus exit.

[00:20:54]Both have the same length, eight and eight plus the exit.

[00:21:00]Great. So, there you have it. 2.6 compared to 2.5 sinking. And as you see, beautiful.