MiniMax M2 <think> tags

Question

Does opencode support rendering tags in the UI like for example Claude Code?

Opencode:

Claude Code:

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

MiniMax-M2 reasoning content not properly handled in OpenAI-compatible mode, should we support this ? #3555: Deals specifically with MiniMax-M2 reasoning content and tags not being properly handled in OpenAI-compatible mode (closed)
[Feature Request] Add support for using thinking tags in local models #1528: General feature request for better support of thinking tags in local models with UI separation

Feel free to ignore if none of these address your specific case.

I'm using Minimax M2 on ~~both their official API and~~ locally hosted with llama.cpp. Both rendering the reasoning blocks properly in opencode, no think tags appear in the response texts. I set reasoning true in opencode.json

"models": {
        "minimax-m2": {
          "name": "minimax-m2",
          "tool_call": true,
          "reasoning": true,
          "cost": {
            "input": 0.10,
            "output": 1.20
          },
          "options": {
            "num_ctx": 131072,
            "temperature": 1.0,
            "top_p": 0.95,
            "top_k": 40
          }
        },

CORRECTION: I am using OpenCode with Minimax-M2 locally hosted only, not the official API. Using Minimax M2 official API with Claude Code.

I'm using offical antropic API, opencode can't get correct title, but no "think" tags.

meet same error in unoffical codex relay... anyone can help?

what is codex relay?

what is codex relay?

Oh, it means third-party intermediary website. It work well with the Codex CLI, but encountered some issues in OpenCode (can't render think blocks).

@aaronnewsome

I'm using Minimax M2 on both their official API and locally hosted with llama.cpp. Both rendering the reasoning blocks properly in opencode, no think tags appear in the response texts. I set reasoning true in opencode.json
"models": {
        "minimax-m2": {
          "name": "minimax-m2",
          "tool_call": true,
          "reasoning": true,
          "cost": {
            "input": 0.10,
            "output": 1.20
          },
          "options": {
            "num_ctx": 131072,
            "temperature": 1.0,
            "top_p": 0.95,
            "top_k": 40
          }
        },
CORRECTION: I am using OpenCode with Minimax-M2 locally hosted only, not the official API. Using Minimax M2 official API with Claude Code.

@aaronnewsome Does tool calling work with OpenCode + local MiniMax-M2? I tested them but OpenCode doesn't parse/understand MiniMax's way of tool call (XML)

I use OpenCode heavily with Minimax-M2 hosted locally. Over many weeks of heavy daily usage, I see no major issues and it runs very stable, much better than GLM 4.5. When someone says it doesn't work, my first two questions are, what are you serving it with and what chat template are you using. Details of my setup are here

ggml-org/llama.cpp#16932 (comment)

I'm using Minmax through nVidia NIM and I can see the tags showing as <think> and I think I saw them as _think_ one time. Seems to me there is a escaping/encode issue with some providers.

V1.0.134 - Windows 10 - Windows Terminal

Same issue. I suspect providers like OpenRouter run extra post-processing & parsing layers behind the scenes, which is why their outputs stay intact - unlike what I see when running the same model locally vLLM

the think tags are being sent as assistant messages not reasoning thats why opencode renders them as such

I'm trying to use MiniMax-M2.1 via the Fireworks API (using the default opencode auth setup) and running into this same issue. And not only are the <think> tags coming through as plaintext, but so are <minimax:tool_call> tags, which renders the model unusable.

AFAIK Fireworks does not have any equivalent to the reasoning_split flag that the official MiniMax API supports to ensure OpenAI SDK compatibility.

If feasible, it'd be awesome if OpenCode offered special support interleaved thinking models via the OpenAI-compatible API. The Fireworks-hosted M2.1 deployment is super appealing as a smart, inexpensive, high-TPS model, and I would love to not have to wait on Fireworks (or any other competing model provider) to figure out how to make these interleaved thinking models truly OpenAI compatible.

@rekram1-node

@rekram1-node I think this issue can be closed

To stay organized issues are automatically closed after 90 days of no activity. If the issue is still relevant please open a new one.

fix: render think tags as reasoning blocks + fix MCP MaxListeners ove…

Navigation Menu

Repository navigation

MiniMax M2 <think> tags #4033

Question

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

MiniMax M2 <think> tags #4033

Description

Question

Activity

github-actions commented on Nov 7, 2025

aaronnewsome commented on Nov 7, 2025

no1wudi commented on Nov 7, 2025

WAY29 commented on Nov 18, 2025

rekram1-node commented on Nov 18, 2025

WAY29 commented on Nov 19, 2025

otarkhan commented on Dec 5, 2025

aaronnewsome commented on Dec 5, 2025

druellan commented on Dec 6, 2025

letsrock85 commented on Dec 25, 2025

rekram1-node commented on Dec 25, 2025

h14h commented on Dec 30, 2025

snipeship commented on Jan 26, 2026

github-actions commented on Mar 29, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions