diff --git a/docs/asset-engine/services.yaml b/docs/asset-engine/services.yaml
index 8635e0e..a494c98 100644
--- a/docs/asset-engine/services.yaml
+++ b/docs/asset-engine/services.yaml
@@ -752,7 +752,7 @@ services:
       Apache-2.0 hybrid diffusion+LLM music generation. Multi-minute lyric-aware
       songs with vocals + instrumentation.
     category: music
-    version: 3
+    version: 4
     host: irv-ml1
     endpoint: http://10.100.79.3:8210/generate
     method: POST
@@ -948,16 +948,23 @@ services:
           Wrapper-side cleanup queued — once the upstream model defaults this,
           the catalog field will become optional or be dropped entirely.
     response:
+      # As of wrapper version that ships with image local/ace-step:v1
+      # post 2026-05-11, /generate streams audio bytes inline with
+      # Content-Type set from the audio_format request field. The
+      # in-container output_path is exposed via X-Output-Path header
+      # for log correlation but is no longer load-bearing.
       type: audio
       mime_from_field: audio_format
-      output_field: output_path
     reproducibility:
       seedable: true
       deterministic: true
       notes: >
         actual_seeds parameter exposed; identical seeds + params = identical audio.
         Local infer-api.py patches upstream's broken 24-arg pipeline signature
-        (was 18 in upstream — caused crashes with audio_duration in `format` slot).
+        (was 18 in upstream — caused crashes with audio_duration in `format` slot)
+        AND inline-streams the generated audio bytes (was returning a JSON
+        path reference to a file inside the container, which was unreachable
+        from outside).
     estimated_latency:
       cold_start_s: 30
       warm_per_unit: "~10–60s depending on audio_duration + infer_step"
diff --git a/stacks/ace-step/infer-api.py b/stacks/ace-step/infer-api.py
index d53624d..4c02bf1 100644
--- a/stacks/ace-step/infer-api.py
+++ b/stacks/ace-step/infer-api.py
@@ -16,6 +16,7 @@
 # audio2audio_enable=False, ref_audio_strength=0.5, no ref audio, no
 # LoRA.)
 from fastapi import FastAPI, HTTPException
+from fastapi.responses import Response
 from pydantic import BaseModel
 from typing import List, Optional
 import os
@@ -24,6 +25,15 @@ import uuid
 
 app = FastAPI(title="ACEStep Pipeline API (patched)")
 
+# audio_format -> Content-Type. The wrapper writes the file in the
+# format the upstream pipeline supports for the requested extension;
+# we just need to label the bytes correctly on the wire.
+_AUDIO_MIME = {
+    "wav":  "audio/wav",
+    "mp3":  "audio/mpeg",
+    "flac": "audio/flac",
+}
+
 
 class ACEStepInput(BaseModel):
     checkpoint_path: str
@@ -78,8 +88,17 @@ def initialize_pipeline(
     )
 
 
-@app.post("/generate", response_model=ACEStepOutput)
+@app.post("/generate")
 async def generate_audio(input_data: ACEStepInput):
+    """Generate music; respond with the audio bytes inline.
+
+    Pre-2026-05-11 this returned a JSON {output_path} reference to a
+    file inside the container — useless to any external consumer
+    since the path wasn't reachable from outside. Now streams the
+    bytes back directly with the right Content-Type, and the file in
+    /app/outputs/ is incidental bookkeeping (the pipeline writes
+    there as part of its normal flow; we read it back for the wire).
+    """
     try:
         model_demo = initialize_pipeline(
             input_data.checkpoint_path,
@@ -115,14 +134,25 @@ async def generate_audio(input_data: ACEStepInput):
             input_data.lora_weight,
         )
 
-        output_path = input_data.output_path or f"/app/outputs/output_{uuid.uuid4().hex}.wav"
+        ext = input_data.audio_format.lower()
+        output_path = (
+            input_data.output_path
+            or f"/app/outputs/output_{uuid.uuid4().hex}.{ext}"
+        )
 
         model_demo(*params, save_path=output_path)
 
-        return ACEStepOutput(
-            status="success",
-            output_path=output_path,
-            message="Audio generated successfully",
+        with open(output_path, "rb") as f:
+            audio_bytes = f.read()
+
+        return Response(
+            content=audio_bytes,
+            media_type=_AUDIO_MIME.get(ext, "application/octet-stream"),
+            headers={
+                # Surfaces the in-container path for debugging/log correlation
+                # without putting it on the response body.
+                "X-Output-Path": output_path,
+            },
         )
 
     except Exception as e: