diff --git a/.env.example b/.env.example
index 3b8a2a5b9534..1e0736921bb4 100644
--- a/.env.example
+++ b/.env.example
@@ -60,6 +60,10 @@
 # RELAY_TIMEOUT=0
 # Relay HTTP 客户端空闲连接超时时间，单位秒，默认跟随 Go 标准库，设置为0表示不限制
 # RELAY_IDLE_CONN_TIMEOUT=90
+# 等待上游返回响应头的超时时间，单位秒，默认 1800，设置为 0 表示不限制。
+# 仅约束「等待响应头」这一段；响应头返回之后的流式传输不受影响。
+# 注意：非流式请求通常要等上游生成完毕才会返回响应头，因此该值需留足余量。
+# RELAY_RESPONSE_HEADER_TIMEOUT=1800
 # 流模式无响应超时时间，单位秒，如果出现空补全可以尝试改为更大值
 # STREAMING_TIMEOUT=300
 
diff --git a/README.md b/README.md
index 91778e659d3f..d05f8e58ca71 100644
--- a/README.md
+++ b/README.md
@@ -327,6 +327,7 @@ docker run --name new-api -d --restart always \
 | `SQL_DSN` | Database connection string | - |
 | `REDIS_CONN_STRING` | Redis connection string | - |
 | `RELAY_IDLE_CONN_TIMEOUT` | Idle keep-alive timeout for relay HTTP clients, seconds. Defaults to Go standard library behavior; set `0` to disable | `90` |
+| `RELAY_RESPONSE_HEADER_TIMEOUT` | How long the relay waits for upstream **response headers**, seconds; set `0` to disable. Only bounds the header wait -- streaming after the headers arrive is unaffected. Note that non-streaming upstreams usually send headers only once generation finishes, so leave headroom | `1800` |
 | `STREAMING_TIMEOUT` | Streaming timeout (seconds) | `300` |
 | `STREAM_SCANNER_MAX_BUFFER_MB` | Max per-line buffer (MB) for the stream scanner; increase when upstream sends huge image/base64 payloads | `64` |
 | `MAX_REQUEST_BODY_MB` | Max request body size (MB, counted **after decompression**; prevents huge requests/zip bombs from exhausting memory). Exceeding it returns `413` | `32` |
diff --git a/common/constants.go b/common/constants.go
index d6b4fb52284c..4e99f5b2f7aa 100644
--- a/common/constants.go
+++ b/common/constants.go
@@ -162,6 +162,16 @@ var BatchUpdateInterval int
 var RelayTimeout int // unit is second
 
 var RelayIdleConnTimeout int // unit is second
+
+// RelayResponseHeaderTimeout limits how long the relay transport waits for the
+// upstream response headers after the request has been fully written.
+// 0 disables it (previous behaviour: wait forever).
+//
+// Note this is NOT the same as RelayTimeout (http.Client.Timeout), which covers
+// the whole response read and therefore breaks legitimate long streaming calls.
+// ResponseHeaderTimeout only bounds the wait for the response headers; once the
+// headers arrive, streaming is unaffected.
+var RelayResponseHeaderTimeout int // unit is second
 var RelayMaxIdleConns int
 var RelayMaxIdleConnsPerHost int
 
diff --git a/common/init.go b/common/init.go
index 4d4c62b27cac..80c6874003c2 100644
--- a/common/init.go
+++ b/common/init.go
@@ -110,6 +110,7 @@ func InitEnv() {
 	BatchUpdateInterval = GetEnvOrDefault("BATCH_UPDATE_INTERVAL", 5)
 	RelayTimeout = GetEnvOrDefault("RELAY_TIMEOUT", 0)
 	RelayIdleConnTimeout = GetEnvOrDefault("RELAY_IDLE_CONN_TIMEOUT", 90)
+	RelayResponseHeaderTimeout = GetEnvOrDefault("RELAY_RESPONSE_HEADER_TIMEOUT", 1800)
 	RelayMaxIdleConns = GetEnvOrDefault("RELAY_MAX_IDLE_CONNS", 500)
 	RelayMaxIdleConnsPerHost = GetEnvOrDefault("RELAY_MAX_IDLE_CONNS_PER_HOST", 100)
 
diff --git a/service/http_client.go b/service/http_client.go
index f5c19eafedc8..19523fb13827 100644
--- a/service/http_client.go
+++ b/service/http_client.go
@@ -4,6 +4,7 @@ import (
 	"context"
 	"crypto/tls"
 	"fmt"
+	"math"
 	"net"
 	"net/http"
 	"net/url"
@@ -71,6 +72,10 @@ func ValidateSSRFProtectedFetchURL(urlStr string) error {
 	return validateURLWithCurrentFetchSetting(urlStr, true)
 }
 
+// maxTimeoutSeconds is the largest number of seconds that still converts to a
+// time.Duration without overflowing (~292 years).
+const maxTimeoutSeconds = int(math.MaxInt64 / int64(time.Second))
+
 func newRelayHTTPTransport() *http.Transport {
 	var transport *http.Transport
 	if defaultTransport, ok := http.DefaultTransport.(*http.Transport); ok && defaultTransport != nil {
@@ -91,6 +96,24 @@ func newRelayHTTPTransport() *http.Transport {
 	transport.MaxIdleConns = common.RelayMaxIdleConns
 	transport.MaxIdleConnsPerHost = common.RelayMaxIdleConnsPerHost
 	transport.IdleConnTimeout = time.Duration(common.RelayIdleConnTimeout) * time.Second
+	// Bound the wait for upstream response headers. Without it, an upstream that
+	// accepts the connection but never responds (and never sends FIN/RST) parks the
+	// goroutine forever, and every buffer that request owns -- the raw body read by
+	// io.ReadAll, the decoded messages, and the re-marshalled upstream body -- stays
+	// reachable for the lifetime of the process.
+	//
+	// This only covers the wait for the headers; streaming after the headers arrive
+	// is not affected. Set RELAY_RESPONSE_HEADER_TIMEOUT=0 to restore the old
+	// unbounded behaviour.
+	if seconds := common.RelayResponseHeaderTimeout; seconds > 0 {
+		// Clamp before converting: seconds beyond maxTimeoutSeconds overflow
+		// time.Duration and can wrap into a tiny positive timeout, which would cut
+		// every relay request instead of only the stuck ones.
+		if seconds > maxTimeoutSeconds {
+			seconds = maxTimeoutSeconds
+		}
+		transport.ResponseHeaderTimeout = time.Duration(seconds) * time.Second
+	}
 	transport.ForceAttemptHTTP2 = true
 	if common.TLSInsecureSkipVerify {
 		transport.TLSClientConfig = common.InsecureTLSConfig