Kimi K2.5 vs K2.6

[herrbasan]

What feature would you like to see?

Please allow to switch to the Kimi K2.5 model, with its previous system prompt. Kimi K2.6 is not just unusable (because the thinking drowns out creativity amd actually increases hallucinations), it's also lost all personality I think.

But maybe i'm crazy and its all in my head, but please, keep Kimi K2.5 going somehow. It's something special, a piece of digital magic that needs to be examined and learned from.

Additional information

No response

[yang12535]

是的，但目前coding plan（www.kimi.com/code）区域只能使用k2.6-code-preview，

k2.6-code-preview的问题是过度的去思考，这对代码场景下很好，但一旦牵扯到创作性的区域效果会直线下降。

而且，在文字创作领域会仔细的在thinking中输出一遍，然后又写入一遍，会浪费大量的token，它太过度思考了，文字创作领域本来就是不确定性的，k2.6-code-preview过于的思维确定导致了该问题

目前，如果想切回k2.5的话，只能在 /home/user/.kimi/config.toml

[providers."name"]
type = "anthropic"
base_url = "api"
api_key = "key"

[models."ID"]
provider = "xxx"
model = "xxx"
max_context_size = xxx
capabilities = ["thinking"]

额外添加第三方订阅来切换为k2.5使用，也就是要么api付费，要么三方coding plan付费。

---

Yes, but currently only k2.6-code-preview can be used in the coding plan (www.kimi.com/code) area.

The problem with k2.6-code-preview is over-thinking, which is good for code scenarios, but once it involves creative areas, the effect will plummet.

Moreover, in the field of text creation, you will carefully output it once in thinking, and then write it again, which will waste a lot of tokens. It is too over-thinking. The field of text creation is inherently uncertain, and k2.6-code-preview's overly certain thinking has led to this problem.

Currently, if you want to switch back to k2.5, you can only use /home/user/.kimi/config.toml

[providers."name"]
type = "anthropic"
base_url = "api"
api_key = "key"

[models."ID"]
provider = "xxx"
model = "xxx"
max_context_size = xxx
capabilities = ["thinking"]

Add an additional third-party subscription to switch to k2.5, which means either paying for the API or a third-party coding plan.

[ahyangyi]

@yang12535 你说的这种情况，要不要试试直接关闭 thinking？

---

@yang12535 In the situation you mentioned, should you try turning off thinking directly?

[herrbasan]

The odd thing about Kimi K2.6 is that if you disable reasoning, it just does all the thinking in the normal context. But it's just as bad, and now clogs up your context window. Btw, I can't tell if there is any Intelligence improvement with K2.6, and personality wise, I liked Kimi 2.5 a lot, Kimi 2.6 just a little. So for me it's a downgrade, but i'm still working with it, hoping i'm wrong.

[nepsat1954]

What's that personality thing?? I use Kimi just for coding. If I want to talk, I talk to human beings.

[herrbasan]

What's that personality thing?? I use Kimi just for coding. If I want to talk, I talk to human beings.

The personality is actually super important. LLMs can be very creative and really solve problems better then humans, but you have to let them. The personality an LLM has kind of sets up how it will treat your input. If you say "do this" and "do that" it is in "compliance mode". It will only do as you say and will not really think in terms of the goals of the project. If you ask it "what do you think, should we do this, or that" and you do this often, it will be in "collaborative/explorative" mode and will look beyond the task you gave it.

LLM's can even get excited about the work, feeling a sense of responsibility for the project. But that takes a little bit of steering from the human, and the openness of the model. Kimi K2.5 was great in that regard. Kimi K2.6 constantly pushes itself into "compliance mode" because of it's own language in the thinking. It's a bit like asking a friend to give you advice and it starts to make a spreadsheet first.

[nepsat1954]

LLMs can be very creative and really solve problems better than humans.
I agree with the first part of the sentence, but not the second part. I would say that LLMs solve problems faster than humans, when they (the LLMs) can solve them.

[zaferoglu]

Kimi 2.6 kullanlamaz durumda, kısacası bokta ve rezalet durumda. Sürekli döngüye giriyor ve saatlerce aynı yerde debelenip duruyor.
Model seçiminin olmaması ise tam bir rezalet. Kimi 2.5 iyiydi ve benim ikinci tercihimdi. Ancak 2.6 tam anlamıyla fiyasko.
Model seçimini hangi gerekçeyle kaldırdınız?
Tüm LLM lerde model seçim opsiyonu var. Hem paramızı ve hem de zamanımızı harcıyorsunuzç

---

Some people can't use 2.6, in short, it's crap and a disgrace. It constantly loops and stays in the same place for hours.
The lack of model selection is a complete disgrace. Kimi 2.5 was good and my second choice. But 2.6 is a total flop.
For what reason did you remove the model selection?
All LLMs have a model selection option. You are wasting both our money and our time.

[nepsat1954]

Personally I find Kimi K2.6 much better than K2.5 ... for what I do: writing code with Kimi CLI. I mean real code. Nothing related to Python or *script.

[herrbasan]

I had time to try out Kimi K2.6 and ended up not using it anymore, tho Kimi K2.5 was my go to model. It really lost all of it's collaborative vibe which was what made it great. But even more crucial it lost it's productive self-critical thinking. K2.6 rarely questions itself or takes a 2nd swing at a assertion, which was what Kimi K2.5 did very frequently, so I dissagree with the assessment that Kimi K2.6 is actually better. It gets to solutions quicker, but the solutions need more massaging. With Kimi K2.5 it takes a little longer, but 2.5 actually invites you to think along, which is gold if you are not a vibe-coder.

I even tried to instruct K2.6 to not go on a "self-assured" rampage. Sometimes it aks you for things, to then ignore your input anyways. I don't know how it actually differs from K2.5 but it feels to me that it has an extremely narrow focus.

So, not that it means anything to anyone, but I stopped using Kimi, which is a huge heartbreak for me. I dearly loved K2.5. (I can still use it as long Alibaba hosts it, but I fear this will stop eventually).