diff --git a/Dockerfile b/Dockerfile
index 85ebe4c..cda3483 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -78,11 +78,6 @@ RUN chmod +x /usr/local/bin/nats-server
 # Include OpenAI Whisper license text for runtime-downloaded ONNX artifacts.
 COPY --from=app-builder /app/compute/core/src/whisper/assets/LICENSE.txt /licenses/openai-whisper-LICENSE.txt
 
-# Bind the worker's health endpoint before downloading ONNX models, so cold
-# starts don't exceed the entrypoint's 30s /health/ready gate. Models load
-# lazily on first job instead.
-ENV COMPUTE_PREWARM_MODELS=false
-
 # Expose the port the app runs on
 EXPOSE 3003
 
diff --git a/compute/worker/.env.example b/compute/worker/.env.example
index 15472fa..ebc0ffa 100644
--- a/compute/worker/.env.example
+++ b/compute/worker/.env.example
@@ -39,7 +39,7 @@ S3_ENDPOINT=http://host.docker.internal:8333
 S3_FORCE_PATH_STYLE=true
 
 # Optional tuning
-# COMPUTE_PREWARM_MODELS=true
+# COMPUTE_PREWARM_MODELS=false
 # COMPUTE_JOB_CONCURRENCY=1
 # COMPUTE_WHISPER_TIMEOUT_MS=30000
 # COMPUTE_PDF_TIMEOUT_MS=300000
diff --git a/compute/worker/docker-compose.yml b/compute/worker/docker-compose.yml
index dde5a73..1bbf550 100644
--- a/compute/worker/docker-compose.yml
+++ b/compute/worker/docker-compose.yml
@@ -24,7 +24,7 @@ services:
       PORT: ${PORT:-8081}
       COMPUTE_WORKER_TOKEN: ${COMPUTE_WORKER_TOKEN:-local-compute-token}
       S3_PREFIX: ${S3_PREFIX:-openreader}
-      COMPUTE_PREWARM_MODELS: ${COMPUTE_PREWARM_MODELS:-true}
+      COMPUTE_PREWARM_MODELS: ${COMPUTE_PREWARM_MODELS:-false}
     ports:
       - "8081:8081"
     develop:
diff --git a/compute/worker/src/runtime.ts b/compute/worker/src/runtime.ts
index 1804855..2a43774 100644
--- a/compute/worker/src/runtime.ts
+++ b/compute/worker/src/runtime.ts
@@ -510,7 +510,7 @@ export async function createComputeWorkerApp(options: CreateComputeWorkerAppOpti
   const pdfTimeoutMs = timeoutConfig.pdfTimeoutMs;
   const pdfHardCapMs = timeoutConfig.pdfHardCapMs;
   const pdfAttempts = readIntEnv('COMPUTE_PDF_JOB_ATTEMPTS', 1);
-  const prewarmModels = parseBoolEnv('COMPUTE_PREWARM_MODELS', true);
+  const prewarmModels = parseBoolEnv('COMPUTE_PREWARM_MODELS', false);
   const jobsStreamMaxBytes = readIntEnv('COMPUTE_JOBS_STREAM_MAX_BYTES', 256 * 1024 * 1024);
   const eventsStreamMaxBytes = readIntEnv('COMPUTE_EVENTS_STREAM_MAX_BYTES', 128 * 1024 * 1024);
   const jobStatesMaxBytes = readIntEnv('COMPUTE_JOB_STATES_MAX_BYTES', 64 * 1024 * 1024);
diff --git a/docs-site/docs/deploy/compute-worker.md b/docs-site/docs/deploy/compute-worker.md
index 260158b..b5ec152 100644
--- a/docs-site/docs/deploy/compute-worker.md
+++ b/docs-site/docs/deploy/compute-worker.md
@@ -51,7 +51,7 @@ Common optional variables:
 - `COMPUTE_WORKER_HOST=0.0.0.0`
 - `PORT=8081` for local/manual runs. Platforms like Railway usually inject `PORT`.
 - `LOG_FORMAT=json` and `COMPUTE_LOG_LEVEL=info`
-- `COMPUTE_PREWARM_MODELS=true`
+- `COMPUTE_PREWARM_MODELS=false` by default. Set it to `true` to pre-download ONNX models during worker startup.
 - `COMPUTE_JOB_CONCURRENCY=1`
 - `COMPUTE_WHISPER_TIMEOUT_MS=30000`
 - `COMPUTE_PDF_TIMEOUT_MS=300000`