diff --git a/internal/llm/vllm.go b/internal/llm/vllm.go
index 1f9f950..97d2919 100644
--- a/internal/llm/vllm.go
+++ b/internal/llm/vllm.go
@@ -28,6 +28,12 @@ type vllmRequest struct {
 	TopP              float64   `json:"top_p"`
 	Stop              []string  `json:"stop,omitempty"`
 	Stream            bool      `json:"stream"`
+	// ChatTemplateKwargs is a vLLM extension. Qwen3-family models reason by
+	// default and prepend a plain-text preamble ("Here's a thinking process:")
+	// ahead of the answer — not a <think> block, so it cannot be stripped after
+	// the fact. Every Petal pass parses a JSON object out of the completion, so
+	// an unsuppressed preamble fails the parse outright.
+	ChatTemplateKwargs map[string]any `json:"chat_template_kwargs,omitempty"`
 }
 
 func (c *VLLMClient) body(req CompletionRequest) vllmRequest {
@@ -40,6 +46,8 @@ func (c *VLLMClient) body(req CompletionRequest) vllmRequest {
 		TopP:              req.TopP,
 		Stop:              req.Stop,
 		Stream:            req.Stream,
+
+		ChatTemplateKwargs: map[string]any{"enable_thinking": false},
 	}
 }