diff --git a/lib/net_mosq.c b/lib/net_mosq.c
index e55be79b8..b44ff7f63 100644
--- a/lib/net_mosq.c
+++ b/lib/net_mosq.c
@@ -279,6 +279,390 @@ int net__socket_shutdown(struct mosquitto *mosq)
 	return rc;
 }
 
+#if !defined(WIN32) && !defined(__QNX__)
+#define HAVE_PSELECT
+#endif
+
+#define timespecsub(tsp, usp, vsp)                                   \
+    do {                                                             \
+        (vsp)->tv_sec  = (tsp)->tv_sec  - (usp)->tv_sec;             \
+        (vsp)->tv_nsec = (tsp)->tv_nsec - (usp)->tv_nsec;            \
+        if ((vsp)->tv_nsec < 0) {                                    \
+            (vsp)->tv_sec--;                                         \
+            (vsp)->tv_nsec += 1000000000L;                           \
+        }                                                            \
+    } while (0)
+
+/*
+ * Happy Eyeballs (RFC 8305)
+ * This implementation races IPv6 candidates sequentially and starts IPv4
+ * after a 250ms delay, or immediately if no IPv6 candidate can be started.
+ * This improves connection latency when IPv6 is broken or slow.
+ */
+#define MOSQ_HE_DELAY_MS 250  /* Happy Eyeballs fallback delay */
+
+/*
+ * Holds resolved bind addresses for IPv4/IPv6.
+ * ai_v4 and ai_v6 are lazily populated only when a candidate of that family
+ * is attempted, to avoid unnecessary getaddrinfo() calls.
+ */
+struct net__bind_info {
+	const char *bind_address;
+	int bind_port;
+	struct addrinfo *ai_v4;
+	struct addrinfo *ai_v6;
+};
+
+/*
+ * Represents a single connection attempt target.
+ * pending:
+ *   0 = not attempted
+ *   1 = connect() in progress (EINPROGRESS)
+ *   2 = connected (winner)
+ */
+#define MOSQ_CON_CAND_SIZE 16
+struct net__conn_candidate {
+	struct addrinfo *ai;
+	mosq_sock_t sock;
+	int family;
+	int pending;
+};
+
+/*
+ * Resolves the bind address for a specific family (IPv4 or IPv6).
+ * This is only used when the caller has explicitly requested a bind address.
+ */
+static int net__resolve_bind_ai(const char *bind_address, int bind_port, int family, struct addrinfo **out)
+{
+	char portstr[8];
+	struct addrinfo hints;
+
+	if(!bind_address || !out){
+		return MOSQ_ERR_INVAL;
+	}
+
+	memset(&hints, 0, sizeof(struct addrinfo));
+	hints.ai_family = family;
+	hints.ai_socktype = SOCK_STREAM;
+	hints.ai_flags = AI_NUMERICSERV;
+
+	snprintf(portstr, sizeof(portstr), "%d", (bind_port > 0 && bind_port < 65535) ? bind_port : 0);
+
+	if(getaddrinfo(bind_address, portstr, &hints, out) != 0){
+		return MOSQ_ERR_EAI;
+	}
+
+	return MOSQ_ERR_SUCCESS;
+}
+
+/* build_bind_info: init bind info + resolve v4/v6 if available */
+static int net__build_bind_info(const char *bind_address, struct net__bind_info *bindinfo)
+{
+	struct addrinfo *tmp = NULL;
+
+	if(!bindinfo || bind_address == NULL){
+		return MOSQ_ERR_INVAL;
+	}
+
+	bindinfo->bind_address = bind_address;
+	bindinfo->bind_port = 0;
+	bindinfo->ai_v4 = NULL;
+	bindinfo->ai_v6 = NULL;
+
+	if (net__resolve_bind_ai(bindinfo->bind_address, bindinfo->bind_port, AF_INET, &tmp) == MOSQ_ERR_SUCCESS) {
+		((struct net__bind_info*)bindinfo)->ai_v4 = tmp;
+	}
+
+	if (net__resolve_bind_ai(bindinfo->bind_address, bindinfo->bind_port, AF_INET6, &tmp) == MOSQ_ERR_SUCCESS) {
+		((struct net__bind_info*)bindinfo)->ai_v6 = tmp;
+	}
+
+	if(!bindinfo->ai_v6 && !bindinfo->ai_v4) {
+		return MOSQ_ERR_EAI;
+	}
+	return MOSQ_ERR_SUCCESS;
+}
+
+/*
+ * Create a non-blocking socket and optionally bind it to a user-specified
+ * local address. This function may return immediately successful or mark the
+ * socket as pending if connect() returns EINPROGRESS.
+ */
+static int net__socket_connect_nb(const struct addrinfo *ai, const struct net__bind_info *bindinfo, mosq_sock_t *sock_out)
+{
+	mosq_sock_t s;
+	const struct addrinfo *bai = NULL;
+
+	if(!ai || !sock_out){
+		return MOSQ_ERR_INVAL;
+	}
+
+	s = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol);
+	if(s == INVALID_SOCKET){
+		*sock_out = INVALID_SOCKET;
+		return MOSQ_ERR_ERRNO;
+	}
+
+	/* Set socket to non-blocking mode. */
+	if(net__socket_nonblock(&s)){
+		COMPAT_CLOSE(s);
+		*sock_out = INVALID_SOCKET;
+		return MOSQ_ERR_ERRNO;
+	}
+
+	/* Bind only if user provided bind_address; skip otherwise */
+	if(bindinfo && bindinfo->bind_address && bindinfo->bind_address[0]){
+		if(ai->ai_family == AF_INET){
+			bai = bindinfo->ai_v4;
+		}else if(ai->ai_family == AF_INET6){
+			bai = bindinfo->ai_v6;
+		}
+
+		if(bai){
+			if(bind(s, bai->ai_addr, (socklen_t)bai->ai_addrlen) != 0){
+				COMPAT_CLOSE(s);
+				*sock_out = INVALID_SOCKET;
+				return MOSQ_ERR_ERRNO;
+			}
+		}
+	}
+
+	if(connect(s, ai->ai_addr, (socklen_t)ai->ai_addrlen) == 0){
+		*sock_out = s;
+		return MOSQ_ERR_SUCCESS;
+	}
+
+	WINDOWS_SET_ERRNO();
+	if(errno != EINPROGRESS && errno != COMPAT_EWOULDBLOCK){
+		COMPAT_CLOSE(s);
+		*sock_out = INVALID_SOCKET;
+		return MOSQ_ERR_ERRNO;
+	}
+
+	*sock_out = s;
+	return MOSQ_ERR_CONN_PENDING;
+}
+
+
+/*
+ * Build ordered candidate list: all IPv6 entries first, then IPv4.
+ * Each candidate contains the resolved sockaddr for connect().
+ * Port is set by using the service string or by mutating ai->ai_addr.
+ */
+static int net__build_candidates(const char *host, uint16_t port, struct net__conn_candidate *cands, int *count, struct addrinfo **addrlist)
+{
+	struct addrinfo hints;
+	struct addrinfo *ai;
+	int i = 0;
+
+	if(!addrlist || !cands || !count) {
+		return MOSQ_ERR_INVAL;
+	}
+
+	memset(&hints, 0, sizeof(struct addrinfo));
+	hints.ai_family = AF_UNSPEC;
+	hints.ai_socktype = SOCK_STREAM;
+
+	if(getaddrinfo(host, NULL, &hints, addrlist) != 0){
+		return MOSQ_ERR_EAI;
+	}
+
+	for(ai = *addrlist; ai && i < MOSQ_CON_CAND_SIZE; ai = ai->ai_next){
+		if(ai->ai_family == AF_INET6){
+			((struct sockaddr_in6 *)ai->ai_addr)->sin6_port = htons(port);
+
+			cands[i].ai = ai;
+			cands[i].sock = INVALID_SOCKET;
+			cands[i].family = AF_INET6;
+			cands[i].pending = 0;
+			i++;
+		}
+	}
+	for(ai = *addrlist; ai && i < MOSQ_CON_CAND_SIZE; ai = ai->ai_next){
+		if(ai->ai_family == AF_INET){
+			((struct sockaddr_in *)ai->ai_addr)->sin_port = htons(port);
+
+			cands[i].ai = ai;
+			cands[i].sock = INVALID_SOCKET;
+			cands[i].family = AF_INET;
+			cands[i].pending = 0;
+			i++;
+		}
+	}
+
+	*count = i;
+	return MOSQ_ERR_SUCCESS;
+}
+
+static int net__start_candidate(int family, struct net__conn_candidate *cands, int cand_count, struct net__bind_info *bindinfo)
+{
+	int rc;
+	int next_index = 0;
+
+	/*
+	 * Start the next candidate of the requested family.
+	 * This function skips immediately failing entries and marks successful
+	 * or EINPROGRESS sockets as pending.
+	 */
+	while (next_index < cand_count) {
+		if (cands[next_index].family == family) {
+			rc = net__socket_connect_nb(cands[next_index].ai, bindinfo, &(cands[next_index].sock));
+			if ((rc == MOSQ_ERR_SUCCESS || rc == MOSQ_ERR_CONN_PENDING) &&
+				cands[next_index].sock != INVALID_SOCKET) {
+				cands[next_index].pending = 1;
+				return MOSQ_ERR_SUCCESS;
+			}
+		}
+		next_index++;
+	}
+
+	return MOSQ_ERR_NO_CONN;
+}
+
+
+/*
+ * Main Happy Eyeballs loop.
+ * Sequentially tries all IPv6 candidates. Once HE delay expires, IPv4
+ * candidates begin. All pending sockets enter a select() race; first to
+ * complete successfully wins and all others are closed.
+ */
+static int net__connect_happy_eyeballs(const char *host, uint16_t port, mosq_sock_t *sock, const char *bind_address)
+{
+	int rc = MOSQ_ERR_SUCCESS;
+
+	struct addrinfo *ainfo = NULL;
+	struct net__bind_info bindinfo;
+	struct net__conn_candidate cands[MOSQ_CON_CAND_SIZE];
+	int cand_count = 0;
+
+	fd_set wfds;
+	mosq_sock_t maxfd;
+	int active;
+
+	int i;
+#ifdef HAVE_PSELECT
+	struct timespec local_timeout;
+#else
+	struct timeval local_timeout;
+#endif
+	struct timespec start, end, diff;
+	long elapsed, wait_ms;
+	int v4_started = 0, v6_started = 0;
+
+	int err = 0;
+	socklen_t len;
+	*sock = INVALID_SOCKET;
+
+	rc = net__build_candidates(host, port, cands, &cand_count, &ainfo);
+	if(rc != MOSQ_ERR_SUCCESS || cand_count == 0) {
+		if (ainfo) {
+			freeaddrinfo(ainfo);
+		}
+		return MOSQ_ERR_NO_CONN;
+	}
+
+	memset(&bindinfo, 0, sizeof(struct net__bind_info));
+	net__build_bind_info(bind_address, &bindinfo);
+
+	/* Start first IPv6 candidate immediately (preferred per RFC 8305). */
+	rc = net__start_candidate(AF_INET6, cands, cand_count, &bindinfo);
+	if (rc == MOSQ_ERR_SUCCESS) {
+		v6_started = 1;
+	}
+
+	while(*sock == INVALID_SOCKET){
+		FD_ZERO(&wfds);
+		maxfd = 0;
+		active = 0;
+
+		clock_gettime(CLOCK_MONOTONIC, &end);
+		timespecsub(&end, &start, &diff);
+		elapsed = diff.tv_sec * 1000 + diff.tv_nsec / 1000000L;
+		/*
+		 * Start IPv4 once the fallback delay has elapsed, or immediately if no
+		 * IPv6 candidate could be started.
+		 */
+		if(!v4_started && (elapsed >= MOSQ_HE_DELAY_MS || !v6_started)){
+			net__start_candidate(AF_INET, cands, cand_count, &bindinfo);
+			v4_started = 1;
+		}
+		/* Add all currently pending sockets to fdset for connect completion. */
+		for(i = 0; i < cand_count; i++){
+			if(cands[i].pending &&
+				cands[i].sock != INVALID_SOCKET){
+				FD_SET(cands[i].sock, &wfds);
+				if(cands[i].sock > maxfd) {
+					maxfd = cands[i].sock;
+				}
+				active++;
+			}
+		}
+
+		if(!active) {
+			rc = MOSQ_ERR_NO_CONN;
+			break;
+		}
+
+		wait_ms = (v4_started) ? MOSQ_HE_DELAY_MS : MOSQ_HE_DELAY_MS - elapsed;
+
+		local_timeout.tv_sec = 0;
+#ifdef HAVE_PSELECT
+		local_timeout.tv_nsec = wait_ms * 1000000L;
+		rc = pselect(maxfd + 1, NULL, &wfds, NULL, &local_timeout, NULL);
+#else
+		local_timeout.tv_usec = wait_ms * 1000;
+		rc = select(maxfd + 1, NULL, &wfds, NULL, &local_timeout);
+#endif
+		if(rc <= 0 && v4_started) {
+			rc = MOSQ_ERR_ERRNO;
+			break;
+		}
+
+		for(i = 0; i < cand_count; i++){
+			if(cands[i].pending && cands[i].sock != INVALID_SOCKET &&
+				FD_ISSET(cands[i].sock, &wfds)){
+
+				/*
+				 * Validate the non-blocking connect completion.
+				 * A socket becoming writable does NOT guarantee success;
+				 * SO_ERROR must be checked to detect ECONNREFUSED, ETIMEDOUT, etc.
+				 */
+				if(!getsockopt(cands[i].sock, SOL_SOCKET, SO_ERROR, (char *)&err, &len)){
+					if (err != 0) {
+						COMPAT_CLOSE(cands[i].sock);
+						cands[i].sock = INVALID_SOCKET;
+						cands[i].pending = 0;
+						continue;
+					}
+				}
+				*sock = cands[i].sock;
+				cands[i].pending = 2;
+				rc = MOSQ_ERR_SUCCESS;
+				break;
+			}
+		}
+	}
+	/* Close all sockets except the winner (pending == 2). */
+	for(i = 0; i < cand_count; i++){
+		if(cands[i].pending != 2 && cands[i].sock != INVALID_SOCKET){
+			COMPAT_CLOSE(cands[i].sock);
+		}
+	}
+	/* Free bind-info and address resolution data. */
+	if(bindinfo.ai_v4) {
+		freeaddrinfo(bindinfo.ai_v4);
+	}
+	if(bindinfo.ai_v6) {
+		freeaddrinfo(bindinfo.ai_v6);
+	}
+	freeaddrinfo(ainfo);
+	/*
+	 * Return MOSQ_ERR_SUCCESS if any socket connected,
+	 * otherwise propagate the failure code determined above.
+	 */
+	return rc;
+}
+
 
 #ifdef FINAL_WITH_TLS_PSK
 
@@ -564,6 +948,8 @@ int net__try_connect(const char *host, uint16_t port, mosq_sock_t *sock, const c
 #else
 		return MOSQ_ERR_NOT_SUPPORTED;
 #endif
+	}else if (!blocking){
+		return net__connect_happy_eyeballs(host, port, sock, bind_address);
 	}else{
 		return net__try_connect_tcp(host, port, sock, bind_address, blocking);
 	}