diff --git a/README.md b/README.md
index c192142..06e3a5e 100644
--- a/README.md
+++ b/README.md
@@ -129,6 +129,18 @@ streams queue rather than fail) or a smaller `--ctx`.
 > [`scripts/monitor.sh`](scripts/monitor.sh) (OOM auto-kill guard) the first time
 > at any new `gpu-mem` or `ctx`.
 
+### Startup time
+
+Time to READY is **~3 min** (vs ~12 min on the default loader). `--load-format
+fastsafetensors` (default) reads the 62–71 GiB **straight to the device** —
+**~32 s** vs ~8 min for the default mmap-backed per-tensor read + redundant
+CPU→GPU copy, which is pathologically slow on GB10's current kernel (the copy is
+also physically pointless on unified memory). It falls back to `nogds`
+automatically — no GPUDirect Storage hardware required; override with
+`LOAD_FORMAT=auto`. The remaining ~2.5 min is `torch.compile` (~42 s) + CUDA-graph
+capture (~2 min); the compile cache is persisted at `$HF_HOME/.vllm_cache`, so
+boots after the first skip the compile.
+
 ## Profiles
 
 Selected with `--profile`:
diff --git a/install.sh b/install.sh
index 8e70aef..7d24cbe 100755
--- a/install.sh
+++ b/install.sh
@@ -288,7 +288,7 @@ start_server() {
     docker rm -f "$NAME" >/dev/null 2>&1 || true
     # shellcheck disable=SC2086
     docker run -d --name "$NAME" --gpus all --net=host --ipc=host --ulimit memlock=-1:-1 \
-        -e HF_HOME=/hf -e MAX_MODEL_LEN="$CTX" -e GPU_MEM="$GPU_MEM" \
+        -e HF_HOME=/hf -e VLLM_CACHE_ROOT=/hf/.vllm_cache -e MAX_MODEL_LEN="$CTX" -e GPU_MEM="$GPU_MEM" \
         -e MAX_NUM_SEQS="$MAX_NUM_SEQS" -e MAX_BATCHED_TOKENS="$MAX_BATCHED_TOKENS" ${HF_TOKEN:+-e HF_TOKEN="$HF_TOKEN"} \
         "${model_env[@]}" \
         -v "$HF_HOME:/hf" -v "$REPO_DIR/runtime:/host:ro" "${mounts[@]}" \
diff --git a/runtime/mtp_serve.sh b/runtime/mtp_serve.sh
index 6f76985..4c01db6 100755
--- a/runtime/mtp_serve.sh
+++ b/runtime/mtp_serve.sh
@@ -11,6 +11,7 @@ MAX_MODEL_LEN="${MAX_MODEL_LEN:-262144}"
 GPU_MEM="${GPU_MEM:-0.82}"
 MAX_NUM_SEQS="${MAX_NUM_SEQS:-3}"
 MAX_BATCHED_TOKENS="${MAX_BATCHED_TOKENS:-8192}"
+LOAD_FORMAT="${LOAD_FORMAT:-fastsafetensors}"
 PORT="${PORT:-8000}"
 echo "[mtp] qwen3_5_mtp — backend=$BACKEND, num_speculative_tokens=$NSPEC, model=$MODEL"
 exec vllm serve "$MODEL" \
@@ -23,5 +24,6 @@ exec vllm serve "$MODEL" \
   --no-enable-prefix-caching \
   --enable-chunked-prefill \
   --trust-remote-code \
+  --load-format "$LOAD_FORMAT" \
   --attention-backend "$BACKEND" \
   --speculative-config "{\"method\":\"qwen3_5_mtp\",\"num_speculative_tokens\":$NSPEC,\"model\":\"$MODEL\"}"