diff --git a/doc/book/reference-manual/configuration.md b/doc/book/reference-manual/configuration.md
index c6dce089..1f583fe6 100644
--- a/doc/book/reference-manual/configuration.md
+++ b/doc/book/reference-manual/configuration.md
@@ -25,7 +25,7 @@ db_engine = "lmdb"
 block_size = "1M"
 block_ram_buffer_max = "256MiB"
 block_max_concurrent_reads = 16
-
+block_max_concurrent_writes_per_request =10
 lmdb_map_size = "1T"
 
 compression_level = 1
@@ -99,6 +99,7 @@ Top-level configuration options, in alphabetical order:
 [`allow_world_readable_secrets`](#allow_world_readable_secrets),
 [`block_max_concurrent_reads`](`block_max_concurrent_reads),
 [`block_ram_buffer_max`](#block_ram_buffer_max),
+[`block_max_concurrent_writes_per_request`](#block_max_concurrent_writes_per_request),
 [`block_size`](#block_size),
 [`bootstrap_peers`](#bootstrap_peers),
 [`compression_level`](#compression_level),
@@ -547,6 +548,14 @@ metric in Prometheus: a non-zero number of such events indicates an I/O
 bottleneck on HDD read speed.
 
 
+#### `block_max_concurrent_writes_per_request` (since  `v2.1.0`) {#block_max_concurrent_writes_per_request}
+
+This parameter is designed to adapt to the concurrent write performance of 
+different storage media.Maximum number of parallel block writes per put request 
+Higher values improve throughput but increase memory usage.
+
+Default: 3, Recommended: 10-30 for NVMe, 3-10 for HDD
+
 #### `lmdb_map_size` {#lmdb_map_size}
 
 This parameters can be used to set the map size used by LMDB,
diff --git a/src/api/s3/put.rs b/src/api/s3/put.rs
index 830a7998..b915f2ec 100644
--- a/src/api/s3/put.rs
+++ b/src/api/s3/put.rs
@@ -39,8 +39,6 @@ use crate::encryption::EncryptionParams;
 use crate::error::*;
 use crate::website::X_AMZ_WEBSITE_REDIRECT_LOCATION;
 
-const PUT_BLOCKS_MAX_PARALLEL: usize = 3;
-
 pub(crate) struct SaveStreamResult {
 	pub(crate) version_uuid: Uuid,
 	pub(crate) version_timestamp: u64,
@@ -493,7 +491,7 @@ pub(crate) async fn read_and_put_blocks<S: Stream<Item = Result<Bytes, Error>> +
 			};
 			let recv_next = async {
 				// If more than a maximum number of writes are in progress, don't add more for now
-				if currently_running >= PUT_BLOCKS_MAX_PARALLEL {
+				if currently_running >= ctx.garage.config.block_max_concurrent_writes_per_request {
 					futures::future::pending().await
 				} else {
 					block_rx3.recv().await
diff --git a/src/util/config.rs b/src/util/config.rs
index e351185f..eb889ebe 100644
--- a/src/util/config.rs
+++ b/src/util/config.rs
@@ -45,6 +45,11 @@ pub struct Config {
 	)]
 	pub block_size: usize,
 
+	/// Maximum number of parallel block writes per PUT request
+	/// Higher values improve throughput but increase memory usage
+	/// Default: 3, Recommended: 10-30 for NVMe, 3-10 for HDD
+	#[serde(default = "default_block_max_concurrent_writes_per_request")]
+	pub block_max_concurrent_writes_per_request: usize,
 	/// Number of replicas. Can be any positive integer, but uneven numbers are more favorable.
 	/// - 1 for single-node clusters, or to disable replication
 	/// - 3 is the recommended and supported setting.
@@ -267,6 +272,9 @@ pub struct KubernetesDiscoveryConfig {
 	pub skip_crd: bool,
 }
 
+pub fn default_block_max_concurrent_writes_per_request() -> usize {
+	3
+}
 /// Read and parse configuration
 pub fn read_config(config_file: PathBuf) -> Result<Config, Error> {
 	let config = std::fs::read_to_string(config_file)?;