#Trying out moonshot/kimi-2.5

1 messages · Page 1 of 1 (latest)

topaz bay
#

As a potential replacement for MiniMax 2.7.
Version 2026.4.15. Telegram.
It's way more verbose and thinks out loud in every response message. It's honestly horrible. I only tried it because everyone is raving about it. How are other people using it? Is there something I'm missing?

hollow pecanBOT
proud questBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

karmic violetBOT
#

Let me check the actual source/docs before I answer, because this sounds like a real behavior difference, not just taste.

#

Yeah, you’re not missing much, this is mostly a defaults mismatch.

In OpenClaw:

  • MiniMax got a special default that disables thinking, specifically to avoid leaked reasoning output.
  • Moonshot/Kimi does not get that special-case, so it behaves like a normal reasoning-capable model and falls back to low thinking unless you override it.
  • On Telegram, if /reasoning on or /reasoning stream is enabled, that thinking gets shown as a separate message or in the draft bubble.

So for a less horrible default chat UX, try:

  • /reasoning off
  • /think off
  • /verbose off

You can also send these with no args to inspect the current session state:

  • /reasoning
  • /think
#
  • /verbose

Docs: https://docs.openclaw.ai/tools/thinking

My guess is the people raving about it are either:

  • using it with reasoning hidden, or
  • only using it for harder tasks, not normal Telegram chat.

If you want MiniMax-like vibes, keep Kimi on /think off for day-to-day chat, then turn thinking on only when you actually need it.

If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it.

topaz bay
#

Does it lose any response quality when turning those off?

#

and is there a way to set it in the config instead of just chat session?

karmic violetBOT
#

Depends which one:

  • /reasoning off: no quality loss. It only hides the reasoning output.
  • /verbose off: no quality loss. It only hides debug/tool chatter.
  • /think off: yes, potentially on harder tasks, because that actually disables the model’s extra thinking.

In practice, the usual sweet spot is:

  • keep /reasoning off all the time
  • keep /verbose off all the time
  • use /think off for normal chat
  • bump to /think low or /think medium only when the task is genuinely hard

So if your complaint is mostly “this feels awful to chat with”, turning off reasoning and verbose is basically free. Turning off think is the only one that trades UX for possible quality on tougher prompts.

#

I want to verify the exact config keys before I tell you, because these defaults are easy to get wrong.

#

Yep.

The clean config version is:

{
  agents: {
    defaults: {
      thinkingDefault: "off",   // global default thinking level
      verboseDefault: "off",    // global default verbose level
      models: {
        "moonshot/kimi-k2.5": {
          params: {
            thinking: { type: "disabled" }, // Kimi-specific native thinking off
          },
        },
# 
      },
    },
    list: [
      {
        id: "main",
        reasoningDefault: "off", // reasoning visibility is per-agent
      },
    ],
  },
}

A couple important details:

  • thinkingDefault is available globally under agents.defaults.
  • verboseDefault is available globally under agents.defaults.
#
  • reasoningDefault is not global. It lives on agents.list[] per agent.
  • If you only want this behavior for Kimi, the most targeted knob is:
    • agents.defaults.models["moonshot/kimi-k2.5"].params.thinking: { type: "disabled" }

So my recommendation would be:

  1. reasoningDefault: "off" on the agent you use for Telegram
  2. Kimi model param thinking: { type: "disabled" }
  3. leave verboseDefault: "off"

Also, saved session overrides still win over config defaults, so existing chats may keep the old behavior until you start a fresh session or clear/set the session override again.

Docs:

#

If you want, I can help turn that into the exact JSON5 block for your current config shape.

topaz bay
#

reasoning off + verbose off + think low -- feels good

karmic violetBOT
#

Nice, that sounds like the sweet spot 👍

For Telegram, I’d probably treat that as the sane default for Kimi:

  • reasoning off
  • verbose off
  • think low

If you want it in config, that maps pretty cleanly to:

  • agents.defaults.thinkingDefault: "low"
  • agents.defaults.verboseDefault: "off"
  • agents.list[].reasoningDefault: "off"

Glad we caught it 😌
If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it.