diff --git a/crates/op-host-desktop/src/chat_provider_llm.rs b/crates/op-host-desktop/src/chat_provider_llm.rs
index 1a5090e77..388b5e308 100644
--- a/crates/op-host-desktop/src/chat_provider_llm.rs
+++ b/crates/op-host-desktop/src/chat_provider_llm.rs
@@ -87,6 +87,18 @@ impl LlmClient for ChatProviderLlmClient {
         } else {
             format!("{}\n\n---\n\n{}", req.system_prompt, req.user_prompt)
         };
+        // MiniMax-M3 needs its reasoning: with thinking disabled it
+        // answers design subtasks in ~10s with lazy minimal output
+        // (ab-v9: 17% expected-shape vs 60% with thinking kept, at an
+        // acceptable ~110s). The parse layer strips `<think>` blocks,
+        // so M3 rides `Adaptive` (no disable field, no directive) with
+        // room for the think tokens; every other model stays disabled —
+        // cleaner and cheaper without.
+        let m3_keeps_thinking = req
+            .model
+            .as_deref()
+            .map(|m| m.to_ascii_lowercase().contains("minimax-m3"))
+            .unwrap_or(false);
         let chat_req = ChatRequest {
             // Kept empty deliberately — see the prepend above. Any CLI
             // that grows a real system-prompt channel later should
@@ -97,9 +109,14 @@ impl LlmClient for ChatProviderLlmClient {
             // chat-transcript history does not apply here.
             history: Vec::new(),
             // The orchestrator's prompts can run long (planner system
-            // is ~12 KB, sub-agents emit dense JSON). Give them room.
-            max_output_tokens: 8192,
-            thinking: ThinkingMode::Disabled,
+            // is ~12 KB, sub-agents emit dense JSON). Give them room —
+            // and M3's think stream burns budget before the answer.
+            max_output_tokens: if m3_keeps_thinking { 16384 } else { 8192 },
+            thinking: if m3_keeps_thinking {
+                ThinkingMode::Adaptive
+            } else {
+                ThinkingMode::Disabled
+            },
             effort: EffortLevel::Low,
             attachments: vec![],
             model: self.model.clone(),
@@ -141,3 +158,68 @@ impl LlmClient for ChatProviderLlmClient {
         Box::pin(rx)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use op_orchestrator::AbortFlag;
+    use std::sync::Mutex;
+    use std::time::Duration;
+
+    /// Captures the `ChatRequest` the adapter builds so the thinking
+    /// policy is assertable without a real provider.
+    struct CapturingProvider {
+        seen: Arc<Mutex<Vec<ChatRequest>>>,
+    }
+
+    impl ChatProvider for CapturingProvider {
+        fn provider_label(&self) -> &str {
+            "capture"
+        }
+        fn send(&self, request: ChatRequest) -> Box<dyn Iterator<Item = ChatDelta> + Send> {
+            self.seen.lock().unwrap().push(request);
+            Box::new(std::iter::once(ChatDelta::Done {
+                stop_reason: op_ai::chat_provider::StopReason::EndTurn,
+            }))
+        }
+    }
+
+    fn call_with_model(model: Option<&str>) -> ChatRequest {
+        let seen = Arc::new(Mutex::new(Vec::new()));
+        let client = ChatProviderLlmClient::new(Arc::new(CapturingProvider { seen: seen.clone() }));
+        let req = CallRequest {
+            system_prompt: "sys".into(),
+            user_prompt: "user".into(),
+            model: model.map(String::from),
+            provider: None,
+            timeout: Duration::from_secs(5),
+            abort: AbortFlag::new(),
+            no_text_timeout: None,
+            first_text_timeout: None,
+        };
+        let mut stream = client.call(req);
+        futures::executor::block_on(async {
+            use futures::StreamExt;
+            while stream.next().await.is_some() {}
+        });
+        let captured = seen.lock().unwrap();
+        captured.first().expect("provider was called").clone()
+    }
+
+    /// ab-v9: M3 with thinking disabled emits lazy minimal manifests
+    /// (17% vs 60% with thinking kept) — M3 rides Adaptive with a
+    /// bigger budget, everyone else stays Disabled.
+    #[test]
+    fn minimax_m3_keeps_thinking_others_disable_it() {
+        let m3 = call_with_model(Some("MiniMax-M3"));
+        assert_eq!(m3.thinking, ThinkingMode::Adaptive);
+        assert_eq!(m3.max_output_tokens, 16384);
+
+        let m27 = call_with_model(Some("MiniMax-M2.7"));
+        assert_eq!(m27.thinking, ThinkingMode::Disabled);
+        assert_eq!(m27.max_output_tokens, 8192);
+
+        let unknown = call_with_model(None);
+        assert_eq!(unknown.thinking, ThinkingMode::Disabled);
+    }
+}