diff --git a/README.md b/README.md
index f4e8533..7d80c50 100644
--- a/README.md
+++ b/README.md
@@ -164,6 +164,8 @@ print(f"\nThroughput: {tps:.2f} tok/s")
 
 All benchmarks share the same datasets (gsm8k, math500, humaneval, mbpp, mt-bench). Datasets are automatically downloaded and cached as JSONL in `cache/` on first run.
 
+Benchmark over a steady request window, not the first requests after launch. The DFlash draft path needs a few requests to warm up; cold readings can undersell steady-state throughput by a large margin (see #135). The harness warms up automatically (`--num-warmup`, default 8); raise it on slower/larger setups.
+
 **vLLM**:
 ```bash
 python -m dflash.benchmark --backend vllm \
diff --git a/dflash/benchmark.py b/dflash/benchmark.py
index 7845a5a..2406901 100644
--- a/dflash/benchmark.py
+++ b/dflash/benchmark.py
@@ -386,7 +386,8 @@ def _run_server(args: argparse.Namespace) -> None:
 
         tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
 
-    num_prompts = args.num_prompts + args.concurrency
+    warmup_n = 0 if args.num_warmup == 0 else max(args.num_warmup, args.concurrency)
+    num_prompts = args.num_prompts + warmup_n
     prompts: list[str] = []
     for i in range(num_prompts):
         item = dataset[i % len(dataset)]
@@ -429,12 +430,11 @@ def send_one(prompt: str) -> dict:
         except Exception:
             print("Warning: /flush_cache failed. Continuing.")
 
-    bs = max(args.concurrency, 1)
-    if len(prompts) > bs:
-        print(f"[warmup] {bs} requests ...")
-        with ThreadPoolExecutor(max_workers=bs) as pool:
-            list(pool.map(send_one, prompts[:bs]))
-        prompts = prompts[bs:]
+    if warmup_n > 0:
+        print(f"[warmup] {warmup_n} requests ...")
+        with ThreadPoolExecutor(max_workers=max(args.concurrency, 1)) as pool:
+            list(pool.map(send_one, prompts[:warmup_n]))
+        prompts = prompts[warmup_n:]
 
     print(f"Running benchmark: {args.num_prompts} prompts, concurrency={args.concurrency} ...")
     start = time.perf_counter()
@@ -492,6 +492,15 @@ def main() -> None:
     parser.add_argument("--base-url", type=str, default="http://127.0.0.1:30000")
     parser.add_argument("--num-prompts", type=int, default=1024)
     parser.add_argument("--concurrency", type=int, default=1)
+    parser.add_argument(
+        "--num-warmup",
+        type=int,
+        default=8,
+        help=(
+            "Warmup requests sent before timing (server backends; default: 8). Draft path needs several requests to "
+            "reach steady state; the documented --concurrency 1 default otherwise warms with only 1 request."
+        ),
+    )
     parser.add_argument("--top-p", type=float, default=1.0)
     parser.add_argument("--top-k", type=int, default=1)
     parser.add_argument("--enable-thinking", action="store_true")