Kimi K2 way better via ACP than via Open AI-compatible route? #41646
bobosola
started this conversation in
LLMs and Zed Agent
Replies: 2 comments 4 replies
-
|
did you find an answer for this. I am still trying to add Moonshot K2 to my ZED but not having any luck. What is the best approach? |
Beta Was this translation helpful? Give feedback.
4 replies
-
|
A I is freaking awesome... so much boilerplate is just handled but still
need to do a log of debugging on large projects. But it gets better every
week! No joke. Estimate it gets you 85% there and then you spend a month
debugging it to end up with a finished product the way you want it. At
least with code bases of 1,000's of lines of code.
All the best,
Jeremy
Yoda of YodaCom
***@***.***
www.yodacom.com
…On Wed, Nov 12, 2025 at 4:20 PM Bob Osola ***@***.***> wrote:
Heh, thanks! I'm late to the AI party but *Zed* and *Kimi K2 Thinking* is
a great combo for the stuff I do without breaking the bank.
—
Reply to this email directly, view it on GitHub
<#41646 (reply in thread)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAMMRQPYVKZ6F4RYYWRWTL334O6FFAVCNFSM6AAAAACKY7CO5SVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIOJVGMYDANY>
.
You are receiving this because you commented.Message ID:
***@***.***>
|
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment

Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
I've just installed the new Kimi CLI which supports ACP and is thus directly callable in Zed in the External Agents list when you create a new AI thread. I also have Kimi K2 installed via the Open AI-compatible route directly to the Moonshot API. Both are using
kimi-k2-0905-preview, and both work great for text and agentic use.But I am seeing much more useful (and very different!) responses when using the model via the ACP interface as opposed to the regular route. I even get very different responses when asking it to enumerate the available tooling. I asked it via Open AI-compatible to suggest improvements to my
main.rs. It did a great job in a couple of minutes and came back with some solid suggestions. But when I asked it to do the same via ACP it really went to town over 20 mins or so and massively re-wrote the code for the better, split upmain.rsinto several more files, added a config.toml, added tests, compiled, corrected bugs, re-compiled, the whole 9 yards.Is there any reason for this huge difference? I'm guessing it's something to do with the Kimi CLI (via ACP) somehow making better use of the model (thinking mode, context handling, settings etc.)? The difference between the two modes is striking. I just wondered if there any general conclusions to be had here regarding using ACP mode over AI-compatible mode?
The only downside of ACP I can see is that there is as yet no history (though I note you can save the responses as Markdown).
Beta Was this translation helpful? Give feedback.
All reactions