diff --git a/dto/gemini.go b/dto/gemini.go
index 489ebea534b4..69b869974a2a 100644
--- a/dto/gemini.go
+++ b/dto/gemini.go
@@ -22,6 +22,97 @@ type GeminiChatRequest struct {
 	CachedContent      string                     `json:"cachedContent,omitempty"`
 }
 
+// GeminiInteractionRequest carries native Vertex interaction payloads such as
+// Lyria 3 requests without forcing them into generateContent semantics.
+type GeminiInteractionRequest struct {
+	Model   *string        `json:"model,omitempty"`
+	Payload map[string]any `json:"-"`
+}
+
+// UnmarshalJSON preserves the complete native interaction payload while
+// exposing model for channel model mapping and billing.
+func (r *GeminiInteractionRequest) UnmarshalJSON(data []byte) error {
+	var payload map[string]any
+	if err := common.Unmarshal(data, &payload); err != nil {
+		return err
+	}
+	r.Payload = payload
+	if model, ok := payload["model"].(string); ok {
+		r.Model = &model
+	}
+	return nil
+}
+
+// MarshalJSON writes the mapped model back into the preserved native payload.
+func (r GeminiInteractionRequest) MarshalJSON() ([]byte, error) {
+	payload := make(map[string]any, len(r.Payload)+1)
+	for k, v := range r.Payload {
+		payload[k] = v
+	}
+	if r.Model != nil {
+		payload["model"] = *r.Model
+	}
+	return common.Marshal(payload)
+}
+
+// GetTokenCountMeta estimates billing text from the native interaction input.
+func (r *GeminiInteractionRequest) GetTokenCountMeta() *types.TokenCountMeta {
+	var inputTexts []string
+	collectGeminiInteractionTexts(r.Payload["input"], &inputTexts)
+	return &types.TokenCountMeta{
+		CombineText: strings.Join(inputTexts, "\n"),
+	}
+}
+
+// IsStream returns false because Vertex interactions are non-streaming.
+func (r *GeminiInteractionRequest) IsStream(c *gin.Context) bool {
+	return false
+}
+
+// SetModelName keeps the URL-selected or mapped model in the native payload.
+func (r *GeminiInteractionRequest) SetModelName(modelName string) {
+	if modelName == "" {
+		return
+	}
+	r.Model = &modelName
+	if r.Payload == nil {
+		r.Payload = map[string]any{}
+	}
+	r.Payload["model"] = modelName
+}
+
+// collectGeminiInteractionTexts extracts text from flexible interaction input
+// arrays without interpreting provider-specific media fields.
+func collectGeminiInteractionTexts(value any, texts *[]string) {
+	switch v := value.(type) {
+	case string:
+		if v != "" {
+			*texts = append(*texts, v)
+		}
+	case []any:
+		for _, item := range v {
+			collectGeminiInteractionTexts(item, texts)
+		}
+	case map[string]any:
+		if itemType, ok := v["type"].(string); ok {
+			if itemType != "text" {
+				return
+			}
+			if text, ok := v["text"].(string); ok && text != "" {
+				*texts = append(*texts, text)
+			}
+			return
+		}
+		if text, ok := v["text"].(string); ok && text != "" {
+			*texts = append(*texts, text)
+			return
+		}
+		for _, item := range v {
+			collectGeminiInteractionTexts(item, texts)
+		}
+	}
+}
+
 // UnmarshalJSON allows GeminiChatRequest to accept both snake_case and camelCase fields.
 func (r *GeminiChatRequest) UnmarshalJSON(data []byte) error {
 	type Alias GeminiChatRequest
@@ -44,9 +135,9 @@ func (r *GeminiChatRequest) UnmarshalJSON(data []byte) error {
 }
 
 type ToolConfig struct {
-	FunctionCallingConfig *FunctionCallingConfig `json:"functionCallingConfig,omitempty"`
-	RetrievalConfig       *RetrievalConfig       `json:"retrievalConfig,omitempty"`
-	IncludeServerSideToolInvocations *bool       `json:"includeServerSideToolInvocations,omitempty"`
+	FunctionCallingConfig            *FunctionCallingConfig `json:"functionCallingConfig,omitempty"`
+	RetrievalConfig                  *RetrievalConfig       `json:"retrievalConfig,omitempty"`
+	IncludeServerSideToolInvocations *bool                  `json:"includeServerSideToolInvocations,omitempty"`
 }
 
 type FunctionCallingConfig struct {
diff --git a/relay/channel/vertex/adaptor.go b/relay/channel/vertex/adaptor.go
index 0d91032d0f35..d02229720454 100644
--- a/relay/channel/vertex/adaptor.go
+++ b/relay/channel/vertex/adaptor.go
@@ -167,6 +167,22 @@ func (a *Adaptor) getRequestUrl(info *relaycommon.RelayInfo, modelName, suffix s
 }
 
 func (a *Adaptor) GetRequestURL(info *relaycommon.RelayInfo) (string, error) {
+	if a.isLyriaInteractionRequest(info) {
+		if info.ChannelOtherSettings.VertexKeyType == dto.VertexKeyTypeAPIKey {
+			return fmt.Sprintf(
+				"%s?key=%s",
+				BuildGoogleModelURL(info.ChannelBaseUrl, OpenSourceAPIVersion, "", "global", info.UpstreamModelName, "interactions"),
+				info.ApiKey,
+			), nil
+		}
+		adc := &Credentials{}
+		if err := common.Unmarshal([]byte(info.ApiKey), adc); err != nil {
+			return "", fmt.Errorf("failed to decode credentials file: %w", err)
+		}
+		a.AccountCredentials = *adc
+		return fmt.Sprintf("%s/interactions", BuildAPIBaseURL(info.ChannelBaseUrl, OpenSourceAPIVersion, adc.ProjectID, "global")), nil
+	}
+
 	suffix := ""
 	if a.RequestMode == RequestModeGemini {
 		if model_setting.GetGeminiSettings().ThinkingAdapterEnabled &&
@@ -328,6 +344,9 @@ func (a *Adaptor) DoRequest(c *gin.Context, info *relaycommon.RelayInfo, request
 }
 
 func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycommon.RelayInfo) (usage any, err *types.NewAPIError) {
+	if a.isLyriaInteractionRequest(info) {
+		return a.doInteractionResponse(c, resp, info)
+	}
 	claudeAdaptor := claude.Adaptor{}
 	if info.IsStream {
 		switch a.RequestMode {
@@ -362,6 +381,54 @@ func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycom
 	return
 }
 
+// isLyriaInteractionRequest identifies the native Vertex interactions endpoint
+// used by Lyria 3 models on the Gemini relay path.
+func (a *Adaptor) isLyriaInteractionRequest(info *relaycommon.RelayInfo) bool {
+	if info == nil {
+		return false
+	}
+	requestPath := strings.Split(info.RequestURLPath, "?")[0]
+	return info.RelayMode == constant.RelayModeGemini &&
+		strings.Contains(requestPath, ":interactions") &&
+		strings.HasPrefix(info.UpstreamModelName, "lyria-3-")
+}
+
+// doInteractionResponse preserves the native Vertex interaction JSON so callers
+// can read all text, metadata, status, and audio outputs.
+func (a *Adaptor) doInteractionResponse(c *gin.Context, resp *http.Response, info *relaycommon.RelayInfo) (*dto.Usage, *types.NewAPIError) {
+	defer resp.Body.Close()
+	bodyBytes, readErr := io.ReadAll(resp.Body)
+	if readErr != nil {
+		return nil, types.NewErrorWithStatusCode(readErr, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError)
+	}
+	bodySnippet := string(bodyBytes)
+	if len(bodySnippet) > 512 {
+		bodySnippet = bodySnippet[:512]
+	}
+	if resp.StatusCode < http.StatusOK || resp.StatusCode >= http.StatusMultipleChoices {
+		return nil, types.NewErrorWithStatusCode(
+			fmt.Errorf("vertex upstream status=%d body=%s", resp.StatusCode, bodySnippet),
+			types.ErrorCodeBadResponseBody,
+			resp.StatusCode,
+		)
+	}
+
+	modelName := strings.TrimSpace(info.UpstreamModelName)
+	if strings.HasPrefix(modelName, "lyria-3-") {
+		contentType := resp.Header.Get("Content-Type")
+		if strings.TrimSpace(contentType) == "" {
+			contentType = "application/json"
+		}
+		c.Data(http.StatusOK, contentType, bodyBytes)
+		return &dto.Usage{
+			PromptTokens: 1,
+			TotalTokens:  1,
+		}, nil
+	}
+
+	return nil, types.NewErrorWithStatusCode(fmt.Errorf("unsupported vertex interaction model: %s", modelName), types.ErrorCodeInvalidRequest, http.StatusBadRequest)
+}
+
 func (a *Adaptor) GetModelList() []string {
 	var modelList []string
 	for i, s := range ModelList {
diff --git a/relay/gemini_handler.go b/relay/gemini_handler.go
index 3b4bafe2a673..c22b87f74be5 100644
--- a/relay/gemini_handler.go
+++ b/relay/gemini_handler.go
@@ -55,6 +55,10 @@ func trimModelThinking(modelName string) string {
 func GeminiHelper(c *gin.Context, info *relaycommon.RelayInfo) (newAPIError *types.NewAPIError) {
 	info.InitChannelMeta(c)
 
+	if interactionReq, ok := info.Request.(*dto.GeminiInteractionRequest); ok {
+		return GeminiInteractionHelper(c, info, interactionReq)
+	}
+
 	geminiReq, ok := info.Request.(*dto.GeminiChatRequest)
 	if !ok {
 		return types.NewErrorWithStatusCode(fmt.Errorf("invalid request type, expected *dto.GeminiChatRequest, got %T", info.Request), types.ErrorCodeInvalidRequest, http.StatusBadRequest, types.ErrOptionWithSkipRetry())
@@ -198,6 +202,66 @@ func GeminiHelper(c *gin.Context, info *relaycommon.RelayInfo) (newAPIError *typ
 	return nil
 }
 
+// GeminiInteractionHelper relays native Gemini/Vertex interaction payloads
+// without converting them into generateContent requests.
+func GeminiInteractionHelper(c *gin.Context, info *relaycommon.RelayInfo, request *dto.GeminiInteractionRequest) (newAPIError *types.NewAPIError) {
+	if !strings.HasPrefix(info.OriginModelName, "lyria-3-") {
+		return types.NewErrorWithStatusCode(fmt.Errorf("unsupported gemini interaction model: %s", info.OriginModelName), types.ErrorCodeInvalidRequest, http.StatusBadRequest, types.ErrOptionWithSkipRetry())
+	}
+
+	err := helper.ModelMappedHelper(c, info, request)
+	if err != nil {
+		return types.NewError(err, types.ErrorCodeChannelModelMappedError, types.ErrOptionWithSkipRetry())
+	}
+
+	adaptor := GetAdaptor(info.ApiType)
+	if adaptor == nil {
+		return types.NewError(fmt.Errorf("invalid api type: %d", info.ApiType), types.ErrorCodeInvalidApiType, types.ErrOptionWithSkipRetry())
+	}
+
+	adaptor.Init(info)
+
+	jsonData, err := common.Marshal(request)
+	if err != nil {
+		return types.NewError(err, types.ErrorCodeConvertRequestFailed, types.ErrOptionWithSkipRetry())
+	}
+
+	if len(info.ParamOverride) > 0 {
+		jsonData, err = relaycommon.ApplyParamOverrideWithRelayInfo(jsonData, info)
+		if err != nil {
+			return newAPIErrorFromParamOverride(err)
+		}
+	}
+
+	logger.LogDebug(c, "Gemini interaction request body: "+string(jsonData))
+
+	resp, err := adaptor.DoRequest(c, info, bytes.NewReader(jsonData))
+	if err != nil {
+		logger.LogError(c, "Do gemini interaction request failed: "+err.Error())
+		return types.NewOpenAIError(err, types.ErrorCodeDoRequestFailed, http.StatusInternalServerError)
+	}
+
+	statusCodeMappingStr := c.GetString("status_code_mapping")
+	httpResp, ok := resp.(*http.Response)
+	if !ok || httpResp == nil {
+		return types.NewErrorWithStatusCode(fmt.Errorf("invalid gemini interaction response type: %T", resp), types.ErrorCodeDoRequestFailed, http.StatusInternalServerError)
+	}
+	if httpResp.StatusCode != http.StatusOK {
+		newAPIError = service.RelayErrorHandler(c.Request.Context(), httpResp, false)
+		service.ResetStatusCode(newAPIError, statusCodeMappingStr)
+		return newAPIError
+	}
+
+	usage, openaiErr := adaptor.DoResponse(c, httpResp, info)
+	if openaiErr != nil {
+		service.ResetStatusCode(openaiErr, statusCodeMappingStr)
+		return openaiErr
+	}
+
+	service.PostTextConsumeQuota(c, info, usage.(*dto.Usage), nil)
+	return nil
+}
+
 func GeminiEmbeddingHandler(c *gin.Context, info *relaycommon.RelayInfo) (newAPIError *types.NewAPIError) {
 	info.InitChannelMeta(c)
 
diff --git a/relay/helper/valid_request.go b/relay/helper/valid_request.go
index 2581b2812c94..bebc53f01af0 100644
--- a/relay/helper/valid_request.go
+++ b/relay/helper/valid_request.go
@@ -27,6 +27,8 @@ func GetAndValidateRequest(c *gin.Context, format types.RelayFormat) (request dt
 			request, err = GetAndValidateGeminiEmbeddingRequest(c)
 		} else if strings.Contains(c.Request.URL.Path, ":batchEmbedContents") {
 			request, err = GetAndValidateGeminiBatchEmbeddingRequest(c)
+		} else if strings.Contains(c.Request.URL.Path, ":interactions") {
+			request, err = GetAndValidateGeminiInteractionRequest(c)
 		} else {
 			request, err = GetAndValidateGeminiRequest(c)
 		}
@@ -321,6 +323,28 @@ func GetAndValidateGeminiRequest(c *gin.Context) (*dto.GeminiChatRequest, error)
 	return request, nil
 }
 
+// GetAndValidateGeminiInteractionRequest validates native Vertex interaction
+// payloads used by Lyria 3 on the Gemini relay path.
+func GetAndValidateGeminiInteractionRequest(c *gin.Context) (*dto.GeminiInteractionRequest, error) {
+	request := &dto.GeminiInteractionRequest{}
+	err := common.UnmarshalBodyReusable(c, request)
+	if err != nil {
+		return nil, err
+	}
+	if request.Payload == nil {
+		return nil, errors.New("request body is required")
+	}
+	input, ok := request.Payload["input"]
+	if !ok {
+		return nil, errors.New("input is required")
+	}
+	inputItems, ok := input.([]any)
+	if !ok || len(inputItems) == 0 {
+		return nil, errors.New("input must be a non-empty array")
+	}
+	return request, nil
+}
+
 func GetAndValidateGeminiEmbeddingRequest(c *gin.Context) (*dto.GeminiEmbeddingRequest, error) {
 	request := &dto.GeminiEmbeddingRequest{}
 	err := common.UnmarshalBodyReusable(c, request)
diff --git a/setting/ratio_setting/model_ratio.go b/setting/ratio_setting/model_ratio.go
index 80702ee42ad2..e86ffeaf86c3 100644
--- a/setting/ratio_setting/model_ratio.go
+++ b/setting/ratio_setting/model_ratio.go
@@ -279,6 +279,8 @@ var defaultModelRatio = map[string]float64{
 var defaultModelPrice = map[string]float64{
 	"suno_music":                     0.1,
 	"suno_lyrics":                    0.01,
+	"lyria-3-pro-preview":            0.08,
+	"lyria-3-clip-preview":           0.04,
 	"dall-e-3":                       0.04,
 	"imagen-3.0-generate-002":        0.03,
 	"black-forest-labs/flux-1.1-pro": 0.04,