How I got Qwen3.5 35B A3B to run with Oh My Pi

README.md

This is how I got unsloth/qwen3.5-35b-a3b to work with Oh My Pi.

1. You need to patch the openai-completions.ts file in the Pi source code to handle Qwen's <think> blocks.

2. You need to create a file at ~/.omp/agent/models.yml to configure the thinking levels for the model.

models.yml

providers:
  lm-studio:
    auth: none
    modelOverrides:
      "unsloth/qwen3.5-35b-a3b":
        reasoning: true
        thinking:
          minLevel: minimal
          maxLevel: high
          mode: effort

openai-completions.ts.diff

--- a/pi-ai/src/providers/openai-completions.ts
+++ b/pi-ai/src/providers/openai-completions.ts
@@ -242,7 +242,7 @@
                      };

-                     const parseMiniMaxThinkTags = model.provider === "minimax-code" || model.provider === "minimax-code-cn";
+                     const parseMiniMaxThinkTags = model.provider === "minimax-code" || model.provider === "minimax-code-cn" || model.provider === "lm-studio";
                      let taggedTextBuffer = "";
                      let insideTaggedThinking = false;