feat(models): 注册 Kimi K3 和 Qwen3.8

apps/api/internal/clients/chat_reasoning.go

@@ -78,6 +78,15 @@ func applyAliyunReasoning(body map[string]any, candidate store.RuntimeModelCandi
 
 	body["enable_thinking"] = true
 	model := chatReasoningModelName(body, candidate)
+	if isAliyunQwen38MaxPreview(model) {
+		if budget, ok := positiveIntFromAny(body["thinking_budget_tokens"]); ok {
+			body["thinking_budget"] = budget
+			delete(body, "reasoning_effort")
+			return
+		}
+		body["reasoning_effort"] = qwen38MaxPreviewReasoningEffort(effort)
+		return
+	}
 	if isAliyunHighMaxReasoningModel(model) {
 		body["reasoning_effort"] = highMaxReasoningEffort(effort)
 		return
@@ -200,6 +209,10 @@ func isAliyunHighMaxReasoningModel(model string) bool {
 	return strings.Contains(model, "deepseek-v4") || strings.HasPrefix(model, "glm-")
 }
 
+func isAliyunQwen38MaxPreview(model string) bool {
+	return strings.Contains(model, "qwen3.8-max-preview")
+}
+
 func isAliyunThinkingBudgetModel(model string) bool {
 	return strings.Contains(model, "qwen") || strings.Contains(model, "qwq") || strings.Contains(model, "qvq") || strings.Contains(model, "kimi")
 }
@@ -219,6 +232,17 @@ func highMaxReasoningEffort(effort string) string {
 	return "high"
 }
 
+func qwen38MaxPreviewReasoningEffort(effort string) string {
+	switch effort {
+	case "low", "minimal":
+		return "low"
+	case "medium":
+		return "medium"
+	default:
+		return "xhigh"
+	}
+}
+
 func zhipuReasoningEffort(effort string) string {
 	if effort == "max" {
 		return "xhigh"

apps/api/internal/clients/clients_test.go

@@ -529,6 +529,36 @@ func TestOpenAIClientChatReasoningParamsByProvider(t *testing.T) {
 				}
 			},
 		},
+		{
+			name:     "aliyun qwen 3.8 preview preserves supported effort",
+			provider: "aliyun-bailian-openai",
+			model:    "qwen3.8-max-preview",
+			body: map[string]any{
+				"reasoning_effort": "high",
+			},
+			assertion: func(t *testing.T, body map[string]any) {
+				if body["enable_thinking"] != true || body["reasoning_effort"] != "xhigh" {
+					t.Fatalf("aliyun qwen 3.8 should use xhigh reasoning effort, got %+v", body)
+				}
+			},
+		},
+		{
+			name:     "aliyun qwen 3.8 preview uses explicit budget exclusively",
+			provider: "aliyun-bailian-openai",
+			model:    "qwen3.8-max-preview",
+			body: map[string]any{
+				"reasoning_effort":       "medium",
+				"thinking_budget_tokens": 16384,
+			},
+			assertion: func(t *testing.T, body map[string]any) {
+				if body["enable_thinking"] != true || body["thinking_budget"] != float64(16384) {
+					t.Fatalf("aliyun qwen 3.8 should use the explicit thinking budget, got %+v", body)
+				}
+				if _, ok := body["reasoning_effort"]; ok {
+					t.Fatalf("aliyun qwen 3.8 must not combine reasoning effort with thinking budget: %+v", body)
+				}
+			},
+		},
 		{
 			name:     "same deepseek model on aliyun uses aliyun protocol",
 			provider: "aliyun-bailian-openai",