fix: preserve Qwen thinking_budget passthrough

relay/channel/ali/adaptor.go

@@ -176,7 +176,7 @@ func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayIn
 
 	switch info.RelayMode {
 	default:
-		aliReq := requestOpenAI2Ali(*request)
+		aliReq := requestOpenAI2Ali(*request, info.UpstreamModelName)
 		return aliReq, nil
 	}
 }

relay/channel/ali/adaptor_test.go

@@ -0,0 +1,128 @@
+package ali
+
+import (
+	"encoding/json"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/QuantumNous/new-api/common"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	relayhelper "github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertOpenAIRequestFiltersThinkingBudgetByUpstreamModel(t *testing.T) {
+	tests := []struct {
+		name          string
+		requestModel  string
+		upstreamModel string
+		budget        string
+		wantBudget    bool
+		wantValue     int64
+	}{
+		{
+			name:          "qwen",
+			requestModel:  "qwen-plus",
+			upstreamModel: "qwen-plus",
+			budget:        "128",
+			wantBudget:    true,
+			wantValue:     128,
+		},
+		{
+			name:          "qwq explicit zero",
+			requestModel:  "qwq-32b",
+			upstreamModel: "qwq-32b",
+			budget:        "0",
+			wantBudget:    true,
+			wantValue:     0,
+		},
+		{
+			name:          "unsupported upstream overrides qwen request",
+			requestModel:  "qwen-plus",
+			upstreamModel: "deepseek-r1",
+			budget:        "128",
+			wantBudget:    false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			request := &dto.GeneralOpenAIRequest{
+				Model:          tt.requestModel,
+				EnableThinking: json.RawMessage(`true`),
+				ThinkingBudget: json.RawMessage(tt.budget),
+			}
+			info := &relaycommon.RelayInfo{
+				ChannelMeta: &relaycommon.ChannelMeta{
+					UpstreamModelName: tt.upstreamModel,
+				},
+			}
+
+			convertedValue, err := (&Adaptor{}).ConvertOpenAIRequest(nil, info, request)
+			require.NoError(t, err)
+			converted, ok := convertedValue.(*dto.GeneralOpenAIRequest)
+			require.True(t, ok)
+
+			if tt.wantBudget {
+				assert.Equal(t, tt.budget, string(converted.ThinkingBudget))
+			} else {
+				assert.Nil(t, converted.ThinkingBudget)
+			}
+
+			encoded, err := common.Marshal(converted)
+			require.NoError(t, err)
+
+			assert.True(t, gjson.GetBytes(encoded, "enable_thinking").Bool())
+			value := gjson.GetBytes(encoded, "thinking_budget")
+			assert.Equal(t, tt.wantBudget, value.Exists())
+			if tt.wantBudget {
+				assert.Equal(t, tt.wantValue, value.Int())
+			}
+		})
+	}
+}
+
+func TestConvertOpenAIRequestPreservesExplicitZeroForMappedQwenModel(t *testing.T) {
+	const (
+		clientModel   = "customer-model"
+		upstreamModel = "Qwen/Qwen3-235B-A22B-Thinking-2507"
+	)
+
+	c, _ := gin.CreateTestContext(httptest.NewRecorder())
+	c.Set("model_mapping", `{"customer-model":"Qwen/Qwen3-235B-A22B-Thinking-2507"}`)
+
+	request := &dto.GeneralOpenAIRequest{
+		Model:          clientModel,
+		EnableThinking: json.RawMessage(`true`),
+		ThinkingBudget: json.RawMessage(`0`),
+	}
+	info := &relaycommon.RelayInfo{
+		OriginModelName: clientModel,
+		ChannelMeta: &relaycommon.ChannelMeta{
+			UpstreamModelName: clientModel,
+		},
+	}
+
+	err := relayhelper.ModelMappedHelper(c, info, request)
+	require.NoError(t, err)
+	assert.True(t, info.IsModelMapped)
+	assert.Equal(t, upstreamModel, info.UpstreamModelName)
+	assert.Equal(t, upstreamModel, request.Model)
+
+	convertedValue, err := (&Adaptor{}).ConvertOpenAIRequest(c, info, request)
+	require.NoError(t, err)
+	converted, ok := convertedValue.(*dto.GeneralOpenAIRequest)
+	require.True(t, ok)
+	assert.Equal(t, json.RawMessage(`0`), converted.ThinkingBudget)
+
+	encoded, err := common.Marshal(converted)
+	require.NoError(t, err)
+
+	value := gjson.GetBytes(encoded, "thinking_budget")
+	assert.True(t, value.Exists())
+	assert.Equal(t, int64(0), value.Int())
+}

relay/channel/ali/text.go

@@ -9,7 +9,15 @@ import (
 
 const EnableSearchModelSuffix = "-internet"
 
-func requestOpenAI2Ali(request dto.GeneralOpenAIRequest) *dto.GeneralOpenAIRequest {
+func requestOpenAI2Ali(request dto.GeneralOpenAIRequest, upstreamModelName string) *dto.GeneralOpenAIRequest {
+	modelName := upstreamModelName
+	if modelName == "" {
+		modelName = request.Model
+	}
+	if !dto.IsQwenThinkingBudgetModel(modelName) {
+		request.ThinkingBudget = nil
+	}
+
 	topP := lo.FromPtrOr(request.TopP, 0)
 	if topP >= 1 {
 		request.TopP = lo.ToPtr(0.999)

relay/channel/baidu/adaptor.go

@@ -27,7 +27,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relay/channel/cloudflare/adaptor.go

@@ -28,7 +28,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) Init(info *relaycommon.RelayInfo) {

relay/channel/cohere/adaptor.go

@@ -26,7 +26,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relay/channel/dify/adaptor.go

@@ -33,7 +33,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {
@@ -109,7 +108,6 @@ func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycom
 	} else {
 		return difyHandler(c, info, resp)
 	}
-	return
 }
 
 func (a *Adaptor) GetModelList() []string {

relay/channel/jina/adaptor.go

@@ -28,7 +28,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relay/channel/mistral/adaptor.go

@@ -25,7 +25,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relay/channel/mokaai/adaptor.go

@@ -27,7 +27,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relay/channel/palm/adaptor.go

@@ -26,7 +26,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relay/channel/tencent/adaptor.go

@@ -34,7 +34,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relay/channel/xunfei/adaptor.go

@@ -26,7 +26,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relay/channel/zhipu/adaptor.go

@@ -26,7 +26,6 @@ func (a *Adaptor) ConvertGeminiRequest(*gin.Context, *relaycommon.RelayInfo, *dt
 func (a *Adaptor) ConvertClaudeRequest(*gin.Context, *relaycommon.RelayInfo, *dto.ClaudeRequest) (any, error) {
 	//TODO implement me
 	panic("implement me")
-	return nil, nil
 }
 
 func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {

relaykit/dto/openai_request.go

@@ -89,6 +89,7 @@ type GeneralOpenAIRequest struct {
 	// Ali Qwen Params
 	VlHighResolutionImages json.RawMessage `json:"vl_high_resolution_images,omitempty"`
 	EnableThinking         json.RawMessage `json:"enable_thinking,omitempty"`
+	ThinkingBudget         json.RawMessage `json:"thinking_budget,omitempty"`
 	ChatTemplateKwargs     json.RawMessage `json:"chat_template_kwargs,omitempty"`
 	EnableSearch           json.RawMessage `json:"enable_search,omitempty"`
 	// ollama Params
@@ -107,6 +108,14 @@ type GeneralOpenAIRequest struct {
 	ReasoningSplit json.RawMessage `json:"reasoning_split,omitempty"`
 }
 
+func (r GeneralOpenAIRequest) MarshalJSON() ([]byte, error) {
+	type Alias GeneralOpenAIRequest
+	if !IsQwenThinkingBudgetModel(r.Model) {
+		r.ThinkingBudget = nil
+	}
+	return kitutil.Marshal((*Alias)(&r))
+}
+
 func (r *GeneralOpenAIRequest) GetTokenCountMeta() *types.TokenCountMeta {
 	var tokenCountMeta types.TokenCountMeta
 	var texts = make([]string, 0)
@@ -222,6 +231,14 @@ func IsOpenAIGPT5Model(modelName string) bool {
 	return strings.HasPrefix(modelName, "gpt-5")
 }
 
+func IsQwenThinkingBudgetModel(modelName string) bool {
+	normalized := strings.ToLower(strings.TrimSpace(modelName))
+	return strings.HasPrefix(normalized, "qwen") ||
+		strings.Contains(normalized, "/qwen") ||
+		strings.HasPrefix(normalized, "qwq") ||
+		strings.Contains(normalized, "/qwq")
+}
+
 func (r *GeneralOpenAIRequest) GetSystemRoleName() string {
 	if IsOpenAIReasoningOModel(r.Model) {
 		if !strings.HasPrefix(r.Model, "o1-mini") && !strings.HasPrefix(r.Model, "o1-preview") {
@@ -880,10 +897,19 @@ type OpenAIResponsesRequest struct {
 	ClientMetadata json.RawMessage `json:"client_metadata,omitempty"`
 	// qwen
 	EnableThinking json.RawMessage `json:"enable_thinking,omitempty"`
+	ThinkingBudget json.RawMessage `json:"thinking_budget,omitempty"`
 	// perplexity
 	Preset json.RawMessage `json:"preset,omitempty"`
 }
 
+func (r OpenAIResponsesRequest) MarshalJSON() ([]byte, error) {
+	type Alias OpenAIResponsesRequest
+	if !IsQwenThinkingBudgetModel(r.Model) {
+		r.ThinkingBudget = nil
+	}
+	return kitutil.Marshal((*Alias)(&r))
+}
+
 func (r *OpenAIResponsesRequest) GetTokenCountMeta() *types.TokenCountMeta {
 	var fileMeta = make([]*types.FileMeta, 0)
 	var texts = make([]string, 0)