Minimax M2.5 experience is weird

[mindplay-dk]

Description

Basically, this is the Minimax M2.5 experience with OpenCode:

As you can see, it's not a context limit issue (<10%) and it often gets permanently stuck in these loops.

I'm using the first-party provider/model minimax/fp8 via openrouter:

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "read": "allow",
    "write": "allow",
    "bash": {
      "*": "ask",
      "git *": "deny"
    }
  },
  "agent": {
    // as recommended here:
    // https://huggingface.co/MiniMaxAI/MiniMax-M2.5#inference-parameters
    "build": {
      "temperature": 1.0,
      "top_p": 0.95,
      "top_k": 40,
    },
    "plan": {
      "temperature": 1.0,
      "top_p": 0.95,
      "top_k": 40,
    }
  },
  "provider": {
    "openrouter": {
      "models": {
        "minimax/minimax-m2.5": {
          "name": "Minimax M2.5 (official)",
          "options": {
            "provider": {
              "only": [
                "minimax/fp8"
              ]
            }
          }
        },
      }
    }
  }
}

I suspect the default model parameters aren't suitable, so I tried to apply the officially recommended ones, but I'm not at all confident that these options are being applied and sent - adding them didn't change anything.

How can I tell? opencode debug agent build just shows whatever I put in the JSON file, it doesn't look like anything gets validated.

I have similar problems with Kimi K2.5, also from the first-party provider. I haven't attempted to change it's configuration.

I've been told on the Discord that there may be "something wrong with my setup", but I have no idea what that would even mean? I have OpenCode on 3 different systems (two Windows 11 systems, one Linux) and no plugins or configuration beyond what you saw here, and it's the same everywhere.

Is it possible that openrouter itself is broken?

Kimi and Minimax combined are getting the same amount of traffic on openrouter as Claude Opus:

So it's hard to believe these models are completely useless?

But I'm at my wits end.

I've spent so many hours on this. It's impossible to tell if any config options are actually doing anything.

I wanted to see what the agent is actually doing, but /export only exports the conversation - I can't even see which files it's reading from the project, and the TUI doesn't display this information either. In general, OpenCode is not very generous with this type of information.

Plugins

none

OpenCode version

1.2.14

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

Zorin 18, Windows 11

Terminal

VS Code terminal

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Loop in certain models #3743: Loop in certain models - reports similar looping behavior with MiniMax and Kimi via third-party providers
• Doom loop is not caught when during reasoning or output #12716: Doom loop is not caught when during reasoning or output - similar stuck loop behavior with certain models
• Kimi K2.5 frequently gets in a reiteration loop #14343: Kimi K2.5 frequently gets in a reiteration loop - same symptom of model getting stuck in repetitive loops
• Qwen3 as well as Kimi K2 keep stopping mid-chat #1522: Qwen3 as well as Kimi K2 keep stopping mid-chat via OpenRouter - similar behavior with OpenRouter-routed models

If your issue is distinct (e.g., specifically about config options not being validated/applied, or about the lack of visibility into agent file reads), please clarify that so it can be tracked separately.

[mindplay-dk]

Possible clue here:

https://www.reddit.com/r/LocalLLaMA/comments/1qnblrd/comment/o1shz83/

Same thing happened to me after the latest llama.cpp update - my models went completely bonkers and started giving me responses to questions I never even asked. Had to rollback to fix it, seems like something got broken in the recent commits

Supposedly was fixed here:

ggml-org/llama.cpp#19115

But I guess there's no saying what version of llama.cpp providers use, if in deed that's what they're using...

[artrixdotdev]

Also experiencing this!

[mindplay-dk]

It seems unlikely this issue is caused by OpenCode - there are reports from other users experiencing the same with other tools via OpenRouter.

Since these issues affect both providers on two different platforms (OpenRouter and OpenCode) I have reached out to both Moonshot and MiniMax to try to get to the bottom of this.

[ankushbhatiya]

Same experience. M2.5 going in loop with Opencode.

[watermelon0guy]

Zed with Open Router, same problem. Strangely Open WEB UI is immune to this

[mindplay-dk]

I gave Minimax M2.5 another shot today - first-party provider via openrouter. (EDIT: not the first-party provider because this is still broken. So all random providers... dangit.)

It completed two fairly long sessions - I watched the entire output both times, the longest task ran up to 130K tokens without problems, where previously I had problems already at 10-20K tokens.

I haven't been able to get a response from either OpenRouter or Minimax regarding this issue, on any channel, but I wonder if something has been fixed upstream somewhere?

There's been no update to OpenCode since I had this issue, I don't think anything changed here.

Is anyone else still having these issues?

[kesor]

Had a problem of "duplicate telegram message" happen to me with OpenClaw. Some investigation showed that my setup of OpenClaw -> Cloudflare AI Gateway -> OpenRouter -> MiniMax got resolved with removing OpenRouter from the chain. A plausible cause might be the added delay/latency. In OpenClaw the PR that explains it in full and addresses the issue is in openclaw/openclaw#41662 and it is likely that OpenCode has a similar problem that gets triggered due to latency of the llm responses.

[tgwfc1012-cyber]

I am also experiencing this exact issue when using minimax M2.5 from openrouter and Opencode Go. One pattern i've observed is that it commonly starts this behavior when the context reaches around 20k~50k. I haven't changed the temp of the modal, just default one.

Strangely, Kimi k2.5 with opencode go have no such issue, but the first party provider on open router still has these exact issues. So, either it is provider's issue as described above, or an unique issue from open code.

would like to seek comments maybe from the Opencode Zen Team for the minimax M2.5 provided by them. Haven't got a chance to check out the one through open code Zen, since I heard the model on Go uses a Quantized Version(lite) than the one provided on the pay-as-you-go Zen.