#Think modes

1 messages · Page 1 of 1 (latest)

remote escarp
#

Im primarily using minimax m2.5 and kimi k2.5 for my main agent. Im trying to decide which is better for the main agent. Im on 2025.3.2 version. I noticed that there is now: off, minimal, low, medium, high and adaptive.

Can you explain how adaptive works?

And which setting is best for each model mentioned?

@valid thunder

valid thunderBOT
#

Adaptive in OpenClaw is not “the gateway auto-picks low/medium/high based on the task”. It’s a provider feature:

  • On models that support provider-managed adaptive thinking (currently Anthropic Claude 4.6 family), OpenClaw can ask the model to dynamically allocate its own thinking budget per prompt (vs you forcing a fixed “low/medium/high” effort). Docs: https://docs.openclaw.ai/tools/thinking (and Anthropic’s own “Adaptive thinking” link is in there).

  • On providers/models that don’t support adaptive natively, adaptive effectively behaves like a normal fixed level (in practice it gets treated ~like medium, or just “thinking enabled”, depending on provider).

What to use for your two models

Kimi K2.5 (Moonshot / Kimi)

Moonshot’s “thinking” is binary:

  • /think off → thinking: { type: "disabled" }
  • any non-off level (minimal/low/medium/high/adaptive) → thinking: { type: "enabled" }

So for Kimi K2.5:

  • Default: off (fastest/cheapest)
  • When you need more careful reasoning: set any non-off level (I’d just use low for clarity)
  • Don’t expect a difference between low vs high vs adaptive on Moonshot; it’s basically on/off.
#

Provider docs: https://docs.openclaw.ai/providers/moonshot#native-thinking-mode-moonshot

MiniMax M2.5

MiniMax M2.5 is marked as reasoning: true in OpenClaw’s model metadata, so OpenClaw will treat it as reasoning-capable and (if nothing else is set) tends to default to something like low-effort thinking.

Practical recommendation for MiniMax M2.5:

  • Default: low (good balance for an always-on main agent)
  • Bump to: medium / high when you’re doing harder planning/debugging and don’t mind extra latency/cost
  • adaptive usually won’t buy you anything special here (it’s mainly for Claude 4.6’s adaptive feature), so I wouldn’t pick it for MiniMax unless you’ve tested and prefer it.

Tip: set it per-message vs per-session

  • One message only: /think:high do a deep debug of this…
  • Set session default: send only /think:low (then it sticks for the session)

If you want “which is better as main agent between MiniMax vs Kimi”, that’s more of a benchmarking / preference call; I’d ask in #models with your use-case (coding? chat? tool-use heavy?) because people have up-to-date hands-on comparisons.

remote escarp
valid thunderBOT
#

For MiniMax M2.5: yes — low as your default, and bump to medium/high when you actually need deeper planning/debugging (expect more latency + output tokens).

For Kimi K2.5 (Moonshot): it’s basically two states:

  • off = thinking disabled (fast/cheap)
  • anything else (minimal/low/medium/high/adaptive) = thinking enabled

So on Kimi, “increase to medium/high” usually won’t change behavior vs low — just pick off for normal and low when you want thinking on.

remote escarp
#

I thought they added the additional modes for kimi k2.5 but i normally run it on low anyway.