diff --git a/build.gradle.kts b/build.gradle.kts
index 9d82809..0bd50f4 100644
--- a/build.gradle.kts
+++ b/build.gradle.kts
@@ -25,7 +25,7 @@ repositories {
     mavenCentral()
 }
 
-val openaiClientVersion = "4.0.1"
+val openaiClientVersion = "4.1.0"
 val ktorVersion = "3.0.3"
 val jLatexMathVersion = "1.0.7"
 val commonTextVersion = "1.13.0"
diff --git a/src/main/kotlin/LargeLanguageModels.kt b/src/main/kotlin/LargeLanguageModels.kt
index 378b3a3..55e6d31 100644
--- a/src/main/kotlin/LargeLanguageModels.kt
+++ b/src/main/kotlin/LargeLanguageModels.kt
@@ -1,14 +1,12 @@
 package top.jie65535.mirai
 
-import com.aallam.openai.api.http.Timeout
-import com.aallam.openai.client.Chat
-import com.aallam.openai.client.OpenAI
-import com.aallam.openai.client.OpenAIHost
+import kotlinx.serialization.json.Json
+import kotlinx.serialization.json.JsonObject
+import kotlinx.serialization.json.jsonObject
 import kotlin.time.Duration.Companion.milliseconds
 
 object LargeLanguageModels {
 
-
     /**
      * 系统提示词
      */
@@ -18,46 +16,62 @@ object LargeLanguageModels {
     /**
      * 聊天助手
      */
-    var chat: Chat? = null
+    var chat: ModelService? = null
 
     /**
      * 推理模型
      */
-    var reasoning: Chat? = null
+    var reasoning: ModelService? = null
 
     /**
      * 视觉模型
      */
-    var visual: Chat? = null
+    var visual: ModelService? = null
+
+    private val json = Json {
+        isLenient = true
+        ignoreUnknownKeys = true
+    }
+
+    private fun parseExtraBody(raw: String): JsonObject? {
+        if (raw.isBlank()) return null
+        return try {
+            json.parseToJsonElement(raw).jsonObject
+        } catch (_: Exception) {
+            null
+        }
+    }
 
     fun reload() {
-        // 载入超时时间
         val timeout = PluginConfig.timeout.milliseconds
 
         // 初始化聊天模型
         if (PluginConfig.openAiApi.isNotBlank() && PluginConfig.openAiToken.isNotBlank()) {
-            chat = OpenAI(
+            chat = ModelService(
+                baseUrl = PluginConfig.openAiApi,
                 token = PluginConfig.openAiToken,
-                host = OpenAIHost(baseUrl = PluginConfig.openAiApi),
-                timeout = Timeout(request = timeout, connect = timeout, socket = timeout)
+                timeout = timeout,
+                extraBody = parseExtraBody(PluginConfig.chatModelExtraBody)
             )
         }
 
         // 初始化推理模型
         if (PluginConfig.reasoningModelApi.isNotBlank() && PluginConfig.reasoningModelToken.isNotBlank()) {
-            reasoning = OpenAI(
+            reasoning = ModelService(
+                baseUrl = PluginConfig.reasoningModelApi,
                 token = PluginConfig.reasoningModelToken,
-                host = OpenAIHost(baseUrl = PluginConfig.reasoningModelApi),
-                timeout = Timeout(request = timeout, connect = timeout, socket = timeout)
+                timeout = timeout,
+                extraBody = parseExtraBody(PluginConfig.reasoningModelExtraBody)
             )
         }
 
         // 初始化视觉模型
         if (PluginConfig.visualModelApi.isNotBlank() && PluginConfig.visualModelToken.isNotBlank()) {
-            visual = OpenAI(
+            visual = ModelService(
+                baseUrl = PluginConfig.visualModelApi,
                 token = PluginConfig.visualModelToken,
-                host = OpenAIHost(baseUrl = PluginConfig.visualModelApi),
-                timeout = Timeout(request = timeout, connect = timeout, socket = timeout)
+                timeout = timeout,
+                extraBody = parseExtraBody(PluginConfig.visualModelExtraBody)
             )
         }
 
@@ -78,4 +92,4 @@ object LargeLanguageModels {
             }
         }
     }
-}
\ No newline at end of file
+}
diff --git a/src/main/kotlin/ModelService.kt b/src/main/kotlin/ModelService.kt
new file mode 100644
index 0000000..2508faf
--- /dev/null
+++ b/src/main/kotlin/ModelService.kt
@@ -0,0 +1,88 @@
+package top.jie65535.mirai
+
+import com.aallam.openai.api.chat.ChatCompletionChunk
+import com.aallam.openai.api.chat.ChatCompletionRequest
+import io.ktor.client.*
+import io.ktor.client.call.body
+import io.ktor.client.engine.okhttp.*
+import io.ktor.client.plugins.*
+import io.ktor.client.request.*
+import io.ktor.client.statement.*
+import io.ktor.http.*
+import io.ktor.utils.io.*
+import kotlinx.coroutines.currentCoroutineContext
+import kotlinx.coroutines.flow.Flow
+import kotlinx.coroutines.flow.flow
+import kotlinx.coroutines.isActive
+import kotlinx.serialization.json.*
+import kotlin.time.Duration
+
+class ModelService(
+    val baseUrl: String,
+    val token: String,
+    val timeout: Duration,
+    val extraBody: JsonObject? = null
+) {
+    val httpClient: HttpClient by lazy {
+        HttpClient(OkHttp) {
+            install(HttpTimeout) {
+                val millis = timeout.inWholeMilliseconds
+                requestTimeoutMillis = millis
+                connectTimeoutMillis = millis
+                socketTimeoutMillis = millis
+            }
+            defaultRequest {
+                url(baseUrl)
+                bearerAuth(token)
+            }
+            expectSuccess = true
+        }
+    }
+
+    private val json = Json {
+        isLenient = true
+        ignoreUnknownKeys = true
+        explicitNulls = false
+    }
+
+    fun chatCompletions(request: ChatCompletionRequest): Flow<ChatCompletionChunk> {
+        val requestJson = json.encodeToJsonElement(ChatCompletionRequest.serializer(), request)
+            .jsonObject.toMutableMap()
+        requestJson["stream"] = JsonPrimitive(true)
+        extraBody?.forEach { (key, value) ->
+            requestJson[key] = value
+        }
+        val body = JsonObject(requestJson).toString()
+
+        return flow {
+            httpClient.post("chat/completions") {
+                setBody(body)
+                contentType(ContentType.Application.Json)
+                accept(ContentType.Text.EventStream)
+                headers {
+                    append(HttpHeaders.CacheControl, "no-cache")
+                    append(HttpHeaders.Connection, "keep-alive")
+                }
+            }.let { response ->
+                val channel: ByteReadChannel = response.body()
+                try {
+                    while (currentCoroutineContext().isActive && !channel.isClosedForRead) {
+                        val line = channel.readUTF8Line() ?: continue
+                        when {
+                            line.startsWith("data: [DONE]") -> break
+                            line.startsWith("data: ") -> {
+                                val chunk = json.decodeFromString<ChatCompletionChunk>(
+                                    line.removePrefix("data: ")
+                                )
+                                emit(chunk)
+                            }
+                            else -> continue
+                        }
+                    }
+                } finally {
+                    channel.cancel()
+                }
+            }
+        }
+    }
+}
diff --git a/src/main/kotlin/PluginConfig.kt b/src/main/kotlin/PluginConfig.kt
index bbd04f9..8bd0e6a 100644
--- a/src/main/kotlin/PluginConfig.kt
+++ b/src/main/kotlin/PluginConfig.kt
@@ -38,6 +38,15 @@ object PluginConfig : AutoSavePluginConfig("Config") {
     @ValueDescription("视觉模型")
     var visualModel: String by value("qwen-vl-plus")
 
+    @ValueDescription("聊天模型额外请求体JSON，会合并到请求体中。例如DeepSeek关闭思维: {\"thinking\": {\"type\": \"disabled\"}}")
+    val chatModelExtraBody: String by value("")
+
+    @ValueDescription("推理模型额外请求体JSON，会合并到请求体中。例如DeepSeek启用思维: {\"thinking\": {\"type\": \"enabled\"}}")
+    val reasoningModelExtraBody: String by value("")
+
+    @ValueDescription("视觉模型额外请求体JSON，会合并到请求体中。")
+    val visualModelExtraBody: String by value("")
+
     @ValueDescription("百炼平台API KEY")
     val dashScopeApiKey: String by value("")