Repository navigation
Has anyone been successful at setting up Kimi K2.5 via NanoGPT? #11529
Description
Question
I am astonished to see that some have been successful in using seamlessly Kimi K2.5 (DeepSeek 3.2 too) via NanoGPT in Opencode.
It has not worked a single time since I have countless tried it. It's buggy:
- it loops over in reasoning,
- it fails to call tools,
- it fires sub-agents then forgets about them, only to stop.
Has anyone figured out a setup that works? I am seeking, please, your advice.
Here is what I tried so far:
- Disabling all my plugins (DCP & Oh-my-opencode)
- Disabling the reasoning
- Adding the
"interleaved": { "field": "reasoning_content" }
Shall I just say that Kimi K2.5 works flawlessly via Opencode Zen.
- Opencode version: 1.1.48-1
- OS: Linux
- Plugins:
oh-my-opencode 2.14.0and@tarquinen/opencode-dcp@latest
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon Jan 31, 2026
This issue might be a duplicate of existing issues. Please check:
permissionsField in Agent Frontmatter #11320: [BUG] Kimi K2 RejectspermissionsField in Agent FrontmatterYour description mentions:
These symptoms align with the known issues above, particularly #11357 (JSON Schema validation with empty parameters), #11053 (question tool not working properly), and #11210 (execution stopping).
Feel free to ignore if none of these address your specific case.
rekram1-node commentedon Jan 31, 2026
Can you share your config?
O-ELMA commentedon Feb 1, 2026
@rekram1-node - yes, sir.
Thank you a bunch for your time and effort.
oh-my-opencode.json
opencode.json
dcp.jsonc
Evengard commentedon Feb 7, 2026
Is there any news on that? OpenCode is literally unusable right now with NanoGPT's Kimi K2.5
KusAdama commentedon Mar 15, 2026
Can confirm, in builder mode (TUI) Kimi K2.5/GLM 5 and + Thinking doesn't work at all.
Plan mode sometimes working.
Connected two weeks ago API from: Venice or manually via config the Routster and Kimi K2.5 worked well out of the box with them.
In the Web server there is different
OT: NanoGPT DeepSeek V3.2, V3.1 + Exp Thinking: for any 1st question asked I'm immediately getting back: Your prompt exceeds the model's context length. Please shorten the input and/or reduce max_tokens (there is 0 in context).
For the Hermes 4: Model not supported.
Qwen3 Coder 30B A3B Instruct does work for example.
surfaceflinger commentedon Apr 2, 2026
For me, Kimi K2.5 (both regular and thinking) from nano-gpt initially works and returns stuff, but often fails to do tool calls properly, deeper into work it completely breaks and just hangs (as in, it counts time how long prompt takes, but nothing happens) opencode for 50~ seconds before allowing me to prompt again.
github-actions commentedon Jun 2, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.