diff --git a/frontend/src/api/model/index.ts b/frontend/src/api/model/index.ts
index f5caf233..9a52b855 100644
--- a/frontend/src/api/model/index.ts
+++ b/frontend/src/api/model/index.ts
@@ -1,4 +1,4 @@
-import { get, post, put, del } from '../../utils/request';
+import { get, post, postUpload, put, del } from '../../utils/request';
 import i18n from '@/i18n'
 
 const t = (key: string) => i18n.global.t(key)
@@ -142,6 +142,47 @@ export function deleteModel(id: string): Promise<void> {
   });
 }
 
+export interface ModelDebugOptions {
+  system_prompt?: string
+  temperature?: number
+  top_p?: number
+  max_tokens?: number
+  thinking?: boolean
+}
+
+export interface ModelDebugResult {
+  ok: boolean
+  elapsed_ms: number
+  request: Record<string, unknown>
+  raw_response: unknown
+  observations: Record<string, unknown>
+  error?: string
+}
+
+export async function debugModel(
+  id: string,
+  data: {
+    input?: string
+    documents?: string[]
+    options?: ModelDebugOptions
+    file?: File | null
+  },
+): Promise<ModelDebugResult> {
+  const form = new FormData()
+  form.append('input', data.input || '')
+  form.append('documents', JSON.stringify(data.documents || []))
+  form.append('options', JSON.stringify(data.options || {}))
+  if (data.file) form.append('file', data.file)
+  const response: any = await postUpload(
+    `/api/v1/models/${id}/debug`,
+    form,
+    undefined,
+    { timeout: 300000 },
+  )
+  if (response?.success && response?.data) return response.data
+  throw new Error(response?.message || t('error.model.getFailed'))
+}
+
 // ----------------------------------------------------------------------------
 // Model credential subresource. See mcp-service.ts for the matching MCP API
 // shape and the design notes in internal/handler/dto/mcp.go.
diff --git a/frontend/src/components/ModelDebugDrawer.vue b/frontend/src/components/ModelDebugDrawer.vue
new file mode 100644
index 00000000..844155ba
--- /dev/null
+++ b/frontend/src/components/ModelDebugDrawer.vue
@@ -0,0 +1,697 @@
+<template>
+  <SettingDrawer
+    v-model:visible="drawerVisible"
+    :title="$t('modelSettings.debug.title')"
+    :description="$t('modelSettings.debug.description')"
+    icon="play-circle-stroke"
+    width="560px"
+    :min-width="480"
+    :max-width="900"
+    storage-key="setting-drawer:width:model-debug"
+    :confirm-text="$t('modelSettings.debug.run')"
+    :confirm-loading="running"
+    :confirm-disabled="!canRun"
+    :cancel-text="$t('common.close')"
+    @confirm="runDebug"
+  >
+    <template v-if="result" #footer-left>
+      <t-button variant="outline" @click="copyResult">
+        <template #icon><t-icon name="file-copy" /></template>
+        {{ $t('modelSettings.debug.copyResult') }}
+      </t-button>
+    </template>
+
+    <div class="model-debug">
+      <section class="setting-drawer__section">
+        <h4 class="setting-drawer__section-title">{{ $t('modelSettings.debug.groupModel') }}</h4>
+        <div v-if="availableModelTypes.length > 1" class="form-item">
+          <div class="model-type-options" role="radiogroup" :aria-label="$t('modelSettings.debug.modelType')">
+            <button
+              v-for="option in availableModelTypes"
+              :key="option.value"
+              type="button"
+              class="model-type-option"
+              :class="{ 'is-active': selectedModelType === option.value }"
+              role="radio"
+              :aria-checked="selectedModelType === option.value"
+              @click="selectModelType(option.value)"
+            >
+              <t-icon :name="option.icon" class="model-type-option__icon" />
+              <span class="model-type-option__label">{{ option.label }}</span>
+            </button>
+          </div>
+        </div>
+        <div class="form-item">
+          <label class="form-label">{{ $t('modelSettings.debug.model') }}</label>
+          <t-select
+            v-model="selectedModelId"
+            filterable
+            :placeholder="$t('modelSettings.debug.modelPlaceholder')"
+            :disabled="filteredModels.length === 0"
+            @change="resetResult"
+          >
+            <t-option
+              v-for="model in filteredModels"
+              :key="model.id"
+              :value="model.id!"
+              :label="modelLabel(model)"
+            >
+              <div class="model-option">
+                <span class="model-option__name">{{ modelLabel(model) }}</span>
+                <span class="model-option__meta">{{ vendorLabel(model) }}</span>
+              </div>
+            </t-option>
+          </t-select>
+          <p v-if="filteredModels.length === 0" class="form-desc">
+            {{ $t('modelSettings.debug.noModelsForType') }}
+          </p>
+        </div>
+      </section>
+
+      <template v-if="selectedModel">
+        <section class="setting-drawer__section">
+          <h4 class="setting-drawer__section-title">{{ $t('modelSettings.debug.groupInput') }}</h4>
+          <div v-if="selectedModel.type !== 'ASR'" class="form-item">
+            <label class="form-label">{{ inputLabel }}</label>
+            <t-textarea
+              v-model="input"
+              :placeholder="inputPlaceholder"
+              :autosize="{ minRows: 4, maxRows: 8 }"
+            />
+          </div>
+
+          <div v-if="selectedModel.type === 'Rerank'" class="form-item">
+            <label class="form-label">{{ $t('modelSettings.debug.documents') }}</label>
+            <t-textarea
+              v-model="documentsText"
+              :placeholder="$t('modelSettings.debug.documentsPlaceholder')"
+              :autosize="{ minRows: 4, maxRows: 8 }"
+            />
+            <p class="form-desc">{{ $t('modelSettings.debug.documentsHint') }}</p>
+          </div>
+
+          <div v-if="needsFile" class="form-item">
+            <label class="form-label">{{ fileLabel }}</label>
+            <div class="file-picker">
+              <input
+                ref="fileInputRef"
+                class="file-picker__input"
+                type="file"
+                :accept="selectedModel.type === 'VLLM' ? 'image/*' : 'audio/*'"
+                @change="onNativeFileChange"
+              >
+              <t-button variant="outline" size="small" @click="fileInputRef?.click()">
+                <template #icon><t-icon name="upload" /></template>
+                {{ $t('modelSettings.debug.chooseFile') }}
+              </t-button>
+            </div>
+            <p v-if="file" class="form-desc">{{ file.name }} · {{ formatBytes(file.size) }}</p>
+          </div>
+        </section>
+
+        <section v-if="isChat" class="setting-drawer__section">
+          <h4 class="setting-drawer__section-title">{{ $t('modelSettings.debug.parameters') }}</h4>
+          <div class="parameter-grid">
+            <div class="form-item">
+              <label class="form-label">Temperature</label>
+              <t-input-number v-model="temperature" :min="0" :max="2" :step="0.1" theme="column" />
+            </div>
+            <div class="form-item">
+              <label class="form-label">Top P</label>
+              <t-input-number v-model="topP" :min="0.01" :max="1" :step="0.1" theme="column" />
+            </div>
+            <div class="form-item">
+              <label class="form-label">Max Tokens</label>
+              <t-input-number v-model="maxTokens" :min="1" :max="8192" :step="128" theme="column" />
+            </div>
+          </div>
+          <div class="form-item">
+            <label class="form-label">{{ $t('modelSettings.debug.systemPrompt') }}</label>
+            <t-textarea
+              v-model="systemPrompt"
+              :placeholder="$t('modelSettings.debug.systemPromptPlaceholder')"
+              :autosize="{ minRows: 2, maxRows: 4 }"
+            />
+          </div>
+          <div v-if="supportsThinking" class="form-item">
+            <label class="form-label">{{ $t('modelSettings.debug.thinking') }}</label>
+            <div class="switch-field">
+              <t-switch v-model="thinking" />
+              <span class="form-desc form-desc--inline">{{ $t('modelSettings.debug.thinkingDesc') }}</span>
+            </div>
+          </div>
+        </section>
+
+        <section v-if="result || history.length > 0" class="setting-drawer__section">
+          <h4 class="setting-drawer__section-title">{{ $t('modelSettings.debug.groupResult') }}</h4>
+
+          <div v-if="history.length > 1" class="form-item">
+            <label class="form-label">{{ $t('modelSettings.debug.history') }}</label>
+            <div class="history-list">
+              <button
+                v-for="run in history"
+                :key="run.id"
+                type="button"
+                class="history-item"
+                :class="{ 'history-item--active': result === run.result }"
+                @click="result = run.result"
+              >
+                <span class="history-item__label">{{ run.label }}</span>
+                <span class="history-item__meta">{{ run.result.elapsed_ms }} ms</span>
+              </button>
+            </div>
+          </div>
+
+          <div v-if="result" class="debug-result">
+            <div class="result-banner" :class="result.ok ? 'result-banner--ok' : 'result-banner--error'">
+              <t-icon :name="result.ok ? 'check-circle-filled' : 'close-circle-filled'" />
+              <div class="result-banner__text">
+                <strong>{{ result.ok ? $t('modelSettings.debug.success') : $t('modelSettings.debug.failed') }}</strong>
+                <span>{{ result.elapsed_ms }} ms</span>
+              </div>
+            </div>
+
+            <div v-if="resultMetrics.length > 0" class="metric-list">
+              <span v-for="metric in resultMetrics" :key="metric.key" class="metric-chip">
+                {{ metric.label }}: {{ metric.value }}
+              </span>
+            </div>
+
+            <p v-if="result.error" class="result-error">{{ result.error }}</p>
+
+            <t-tabs v-model="resultTab" class="result-tabs">
+              <t-tab-panel value="response" :label="$t('modelSettings.debug.rawResponse')" />
+              <t-tab-panel value="request" :label="$t('modelSettings.debug.requestPreview')" />
+            </t-tabs>
+            <pre class="json-output">{{ formattedResult }}</pre>
+          </div>
+        </section>
+      </template>
+    </div>
+  </SettingDrawer>
+</template>
+
+<script setup lang="ts">
+import { computed, ref, watch } from 'vue'
+import { MessagePlugin } from 'tdesign-vue-next'
+import { useI18n } from 'vue-i18n'
+import SettingDrawer from '@/components/settings/SettingDrawer.vue'
+import { debugModel, type ModelConfig, type ModelDebugResult } from '@/api/model'
+import { modelSupportsThinking } from '@/utils/thinkingControl'
+
+const props = defineProps<{
+  visible: boolean
+  models: ModelConfig[]
+}>()
+
+const emit = defineEmits<{
+  (e: 'update:visible', value: boolean): void
+}>()
+
+const { t, te } = useI18n()
+const drawerVisible = computed({
+  get: () => props.visible,
+  set: value => emit('update:visible', value),
+})
+
+type DebugModelType = ModelConfig['type']
+
+const selectedModelType = ref<DebugModelType>('KnowledgeQA')
+const selectedModelId = ref('')
+const input = ref('')
+const documentsText = ref('')
+const file = ref<File | null>(null)
+const fileInputRef = ref<HTMLInputElement | null>(null)
+const thinking = ref(false)
+const temperature = ref(0.7)
+const topP = ref(1)
+const maxTokens = ref(1024)
+const systemPrompt = ref('')
+const running = ref(false)
+const result = ref<ModelDebugResult | null>(null)
+const resultTab = ref<'response' | 'request'>('response')
+const history = ref<Array<{
+  id: number
+  label: string
+  result: ModelDebugResult
+}>>([])
+let runSequence = 0
+
+const selectedModel = computed(() => props.models.find(model => model.id === selectedModelId.value))
+const filteredModels = computed(() => props.models.filter(model => model.type === selectedModelType.value))
+const isChat = computed(() => selectedModel.value?.type === 'KnowledgeQA')
+const supportsThinking = computed(() => selectedModel.value ? modelSupportsThinking(selectedModel.value) : false)
+const needsFile = computed(() => ['VLLM', 'ASR'].includes(selectedModel.value?.type || ''))
+const documents = computed(() => documentsText.value.split('\n').map(item => item.trim()).filter(Boolean))
+const canRun = computed(() => {
+  if (!selectedModel.value) return false
+  if (needsFile.value && !file.value) return false
+  if (selectedModel.value.type === 'ASR') return true
+  if (selectedModel.value.type === 'Rerank') return !!input.value.trim() && documents.value.length > 0
+  return !!input.value.trim()
+})
+
+const allModelTypeOptions = computed(() => {
+  const keys: Record<DebugModelType, { short: string; icon: string }> = {
+    KnowledgeQA: { short: 'chat', icon: 'chat' },
+    Embedding: { short: 'embedding', icon: 'chart-bubble' },
+    Rerank: { short: 'rerank', icon: 'filter-sort' },
+    VLLM: { short: 'vllm', icon: 'image' },
+    ASR: { short: 'asr', icon: 'sound' },
+  }
+  return (Object.keys(keys) as DebugModelType[]).map(value => ({
+    value,
+    label: t(`modelSettings.typeShort.${keys[value].short}`),
+    icon: keys[value].icon,
+  }))
+})
+
+const modelCount = (type: DebugModelType) => props.models.filter(model => model.type === type).length
+
+const availableModelTypes = computed(() =>
+  allModelTypeOptions.value.filter(option => modelCount(option.value) > 0),
+)
+
+const modelLabel = (model: ModelConfig) => model.display_name?.trim() || model.name
+
+const vendorLabel = (model: ModelConfig) => {
+  const provider = model.parameters.provider || ''
+  if (model.source === 'local') return 'Ollama'
+  if (provider === 'generic') return t('modelSettings.source.custom')
+  const key = `model.editor.providers.${provider}.label`
+  return te(key) ? t(key) : provider || model.source
+}
+
+const inputLabel = computed(() => {
+  if (selectedModel.value?.type === 'Embedding') return t('modelSettings.debug.embeddingInput')
+  if (selectedModel.value?.type === 'VLLM') return t('modelSettings.debug.vlmPrompt')
+  if (selectedModel.value?.type === 'Rerank') return t('modelSettings.debug.query')
+  return t('modelSettings.debug.query')
+})
+
+const inputPlaceholder = computed(() => {
+  if (selectedModel.value?.type === 'Embedding') return t('modelSettings.debug.embeddingPlaceholder')
+  if (selectedModel.value?.type === 'VLLM') return t('modelSettings.debug.vlmPromptPlaceholder')
+  return t('modelSettings.debug.queryPlaceholder')
+})
+
+const fileLabel = computed(() =>
+  selectedModel.value?.type === 'VLLM'
+    ? t('modelSettings.debug.imageFile')
+    : t('modelSettings.debug.audioFile'),
+)
+
+const formattedResult = computed(() => {
+  if (!result.value) return ''
+  const value = resultTab.value === 'response'
+    ? result.value.raw_response
+    : result.value.request
+  return JSON.stringify(value, null, 2)
+})
+
+const OBSERVATION_LABELS: Record<string, string> = {
+  dimension: 'modelSettings.debug.metrics.dimension',
+  result_count: 'modelSettings.debug.metrics.resultCount',
+  answer_characters: 'modelSettings.debug.metrics.answerChars',
+  reasoning_characters: 'modelSettings.debug.metrics.reasoningChars',
+  reasoning_returned: 'modelSettings.debug.metrics.reasoningReturned',
+  text_characters: 'modelSettings.debug.metrics.textChars',
+  segment_count: 'modelSettings.debug.metrics.segmentCount',
+}
+
+const resultMetrics = computed(() => {
+  if (!result.value?.observations) return []
+  const obs = result.value.observations
+  const keys = Object.keys(OBSERVATION_LABELS).filter(key => obs[key] !== undefined && obs[key] !== null)
+  return keys.map(key => ({
+    key,
+    label: t(OBSERVATION_LABELS[key]),
+    value: formatMetricValue(key, obs[key]),
+  }))
+})
+
+const formatMetricValue = (key: string, value: unknown) => {
+  if (typeof value === 'boolean') {
+    return value ? t('common.yes') : t('common.no')
+  }
+  return String(value)
+}
+
+const ensureDefaultSelection = () => {
+  const types = availableModelTypes.value
+  if (types.length === 0) {
+    selectedModelId.value = ''
+    return
+  }
+  if (!types.some(option => option.value === selectedModelType.value)) {
+    selectedModelType.value = types[0].value
+  }
+  const models = filteredModels.value
+  if (!models.some(model => model.id === selectedModelId.value)) {
+    selectedModelId.value = models[0]?.id || ''
+  }
+}
+
+watch(() => props.visible, visible => {
+  if (visible) ensureDefaultSelection()
+})
+
+watch(availableModelTypes, () => {
+  if (props.visible) ensureDefaultSelection()
+})
+
+watch(() => selectedModel.value?.id, () => {
+  if (!supportsThinking.value) thinking.value = false
+})
+
+watch(() => selectedModel.value?.type, () => {
+  file.value = null
+  result.value = null
+  history.value = []
+  resultTab.value = 'response'
+})
+
+const resetResult = () => {
+  result.value = null
+  history.value = []
+  resultTab.value = 'response'
+}
+
+const selectModelType = (type: DebugModelType) => {
+  if (selectedModelType.value === type) return
+  selectedModelType.value = type
+  selectedModelId.value = filteredModels.value[0]?.id || ''
+  input.value = ''
+  documentsText.value = ''
+  file.value = null
+  resetResult()
+}
+
+const onNativeFileChange = (event: Event) => {
+  const target = event.target as HTMLInputElement
+  file.value = target.files?.[0] || null
+  resetResult()
+}
+
+const formatBytes = (bytes: number) => {
+  if (bytes < 1024) return `${bytes} B`
+  if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(1)} KB`
+  return `${(bytes / (1024 * 1024)).toFixed(1)} MB`
+}
+
+const historyLabel = (thinkingValue: boolean) => {
+  if (supportsThinking.value) {
+    return thinkingValue ? t('modelSettings.debug.thinkOn') : t('modelSettings.debug.thinkOff')
+  }
+  return t('modelSettings.debug.runLabel', { n: runSequence })
+}
+
+const runDebug = async () => {
+  if (!selectedModel.value?.id || !canRun.value || running.value) return
+  running.value = true
+  try {
+    const thinkingValue = supportsThinking.value ? thinking.value : false
+    const nextResult = await debugModel(selectedModel.value.id, {
+      input: input.value.trim(),
+      documents: documents.value,
+      file: file.value,
+      options: isChat.value ? {
+        system_prompt: systemPrompt.value.trim() || undefined,
+        temperature: temperature.value,
+        top_p: topP.value,
+        max_tokens: maxTokens.value,
+        thinking: thinkingValue,
+      } : {},
+    })
+    result.value = nextResult
+    history.value.unshift({
+      id: ++runSequence,
+      label: historyLabel(thinkingValue),
+      result: nextResult,
+    })
+    history.value = history.value.slice(0, 6)
+    resultTab.value = 'response'
+  } catch (error: any) {
+    MessagePlugin.error(error?.message || t('modelSettings.debug.requestFailed'))
+  } finally {
+    running.value = false
+  }
+}
+
+const copyResult = async () => {
+  if (!result.value) return
+  try {
+    await navigator.clipboard.writeText(JSON.stringify(result.value, null, 2))
+    MessagePlugin.success(t('common.copied'))
+  } catch {
+    MessagePlugin.error(t('common.copyFailed'))
+  }
+}
+</script>
+
+<style scoped lang="less">
+.form-item {
+  margin-bottom: 0;
+}
+
+.form-label {
+  display: block;
+  margin-bottom: 6px;
+  font-size: 13px;
+  font-weight: 500;
+  color: var(--td-text-color-primary);
+  line-height: 1.4;
+}
+
+.form-desc {
+  margin: 4px 0 0;
+  font-size: 12px;
+  line-height: 1.5;
+  color: var(--td-text-color-placeholder);
+
+  &--inline {
+    margin: 0;
+  }
+}
+
+.model-type-options {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+}
+
+.model-type-option {
+  display: inline-flex;
+  align-items: center;
+  gap: 6px;
+  padding: 6px 12px;
+  min-height: 32px;
+  border: 1px solid var(--td-component-stroke);
+  border-radius: 8px;
+  background: var(--td-bg-color-container);
+  color: var(--td-text-color-secondary);
+  font-size: 13px;
+  line-height: 1.4;
+  cursor: pointer;
+  transition: border-color 0.15s ease, color 0.15s ease, background 0.15s ease;
+
+  &__icon {
+    font-size: 15px;
+    flex-shrink: 0;
+  }
+
+  &__label {
+    white-space: nowrap;
+  }
+
+  &:hover:not(.is-active) {
+    border-color: var(--td-brand-color-3, var(--td-brand-color));
+    color: var(--td-text-color-primary);
+  }
+
+  &.is-active {
+    border-color: var(--td-brand-color);
+    background: color-mix(in srgb, var(--td-brand-color) 10%, transparent);
+    color: var(--td-brand-color);
+    font-weight: 500;
+  }
+
+  &:focus-visible {
+    outline: 2px solid var(--td-brand-color);
+    outline-offset: 2px;
+  }
+}
+
+.model-option {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 12px;
+  min-width: 0;
+
+  &__name {
+    overflow: hidden;
+    text-overflow: ellipsis;
+    white-space: nowrap;
+  }
+
+  &__meta {
+    flex-shrink: 0;
+    color: var(--td-text-color-placeholder);
+    font-size: 12px;
+  }
+}
+
+.parameter-grid {
+  display: grid;
+  grid-template-columns: repeat(3, minmax(0, 1fr));
+  gap: 14px;
+
+  :deep(.t-input-number) {
+    width: 100%;
+  }
+}
+
+.switch-field {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  min-height: 32px;
+}
+
+.file-picker {
+  position: relative;
+
+  &__input {
+    position: absolute;
+    width: 0;
+    height: 0;
+    opacity: 0;
+    pointer-events: none;
+  }
+}
+
+.history-list {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 6px;
+}
+
+.history-item {
+  display: inline-flex;
+  align-items: center;
+  gap: 8px;
+  padding: 5px 10px;
+  border: 1px solid var(--td-component-stroke);
+  border-radius: 6px;
+  background: var(--td-bg-color-container);
+  color: var(--td-text-color-secondary);
+  cursor: pointer;
+  font: inherit;
+  font-size: 12px;
+
+  &__label {
+    color: var(--td-text-color-primary);
+    font-weight: 500;
+  }
+
+  &__meta {
+    color: var(--td-text-color-placeholder);
+  }
+
+  &:hover,
+  &--active {
+    border-color: var(--td-brand-color);
+    background: color-mix(in srgb, var(--td-brand-color) 6%, var(--td-bg-color-container));
+  }
+}
+
+.result-banner {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+  padding: 10px 12px;
+  border-radius: 8px;
+  font-size: 13px;
+
+  &--ok {
+    background: var(--td-success-color-light);
+    color: var(--td-success-color);
+
+    .result-banner__text strong {
+      color: var(--td-text-color-primary);
+    }
+  }
+
+  &--error {
+    background: var(--td-error-color-light);
+    color: var(--td-error-color);
+
+    .result-banner__text strong {
+      color: var(--td-text-color-primary);
+    }
+  }
+
+  &__text {
+    display: flex;
+    align-items: baseline;
+    gap: 8px;
+
+    span {
+      color: var(--td-text-color-placeholder);
+      font-size: 12px;
+    }
+  }
+}
+
+.metric-list {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 6px;
+  margin-top: 10px;
+}
+
+.metric-chip {
+  padding: 2px 8px;
+  border-radius: 4px;
+  background: var(--td-bg-color-secondarycontainer);
+  color: var(--td-text-color-secondary);
+  font-size: 12px;
+}
+
+.result-error {
+  margin: 10px 0 0;
+  color: var(--td-error-color);
+  font-size: 13px;
+  white-space: pre-wrap;
+}
+
+.result-tabs {
+  margin-top: 12px;
+
+  :deep(.t-tabs__content) {
+    display: none;
+  }
+}
+
+.json-output {
+  overflow: auto;
+  max-height: 420px;
+  min-height: 140px;
+  margin: 8px 0 0;
+  padding: 12px 14px;
+  border: 1px solid var(--td-component-stroke);
+  border-radius: 8px;
+  background: var(--td-bg-color-secondarycontainer);
+  color: var(--td-text-color-primary);
+  font: 12px/1.6 ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, monospace;
+  white-space: pre-wrap;
+  word-break: break-word;
+}
+
+@media (max-width: 640px) {
+  .parameter-grid {
+    grid-template-columns: 1fr;
+  }
+}
+</style>
diff --git a/frontend/src/i18n/locales/en-US.ts b/frontend/src/i18n/locales/en-US.ts
index 05f7ac83..fdefb20d 100755
--- a/frontend/src/i18n/locales/en-US.ts
+++ b/frontend/src/i18n/locales/en-US.ts
@@ -4237,6 +4237,7 @@ export default {
     },
     actions: {
       addModel: 'Add Model',
+      debugModel: 'Model Test',
       setDefault: 'Set as Default'
     },
     source: {
@@ -4298,6 +4299,54 @@ export default {
     },
     builtinTag: 'Built-in',
     confirmDelete: 'Delete model "{name}"?',
+    debug: {
+      title: 'Model Test',
+      description: 'Send a real request to a saved model and inspect the response',
+      groupModel: 'Select model',
+      groupInput: 'Test input',
+      groupResult: 'Result',
+      modelType: 'Model type',
+      model: 'Model',
+      modelPlaceholder: 'Select a model to test',
+      noModelsForType: 'No saved models of this type',
+      query: 'Input',
+      queryPlaceholder: 'Enter the text to send to the model',
+      embeddingInput: 'Text to embed',
+      embeddingPlaceholder: 'Enter text to generate an embedding',
+      vlmPrompt: 'Image prompt',
+      vlmPromptPlaceholder: 'For example: Describe this image in detail',
+      documents: 'Candidate documents',
+      documentsPlaceholder: 'Enter one candidate document per line',
+      documentsHint: 'Each non-empty line is sent as a separate document to the ReRank model',
+      imageFile: 'Image file',
+      audioFile: 'Audio file',
+      chooseFile: 'Choose file',
+      parameters: 'Request parameters',
+      thinking: 'Thinking mode',
+      thinkingDesc: 'Only applies to models that support thinking',
+      systemPrompt: 'System Prompt',
+      systemPromptPlaceholder: 'Optional system prompt',
+      run: 'Run test',
+      copyResult: 'Copy result',
+      history: 'History',
+      thinkOn: 'Thinking on',
+      thinkOff: 'Thinking off',
+      runLabel: 'Run #{n}',
+      success: 'Request succeeded',
+      failed: 'Request failed',
+      rawResponse: 'Response',
+      requestPreview: 'Request preview',
+      requestFailed: 'Model test request failed',
+      metrics: {
+        dimension: 'Dimensions',
+        resultCount: 'Result count',
+        answerChars: 'Answer chars',
+        reasoningChars: 'Reasoning chars',
+        reasoningReturned: 'Reasoning returned',
+        textChars: 'Transcript chars',
+        segmentCount: 'Segment count',
+      },
+    },
   },
   // New: Ollama Settings
   ollamaSettings: {
diff --git a/frontend/src/i18n/locales/ko-KR.ts b/frontend/src/i18n/locales/ko-KR.ts
index f7e201ed..fcdf9f81 100755
--- a/frontend/src/i18n/locales/ko-KR.ts
+++ b/frontend/src/i18n/locales/ko-KR.ts
@@ -4233,6 +4233,7 @@ export default {
     },
     actions: {
       addModel: "모델 추가",
+      debugModel: "모델 테스트",
       setDefault: "기본값으로 설정",
     },
     source: {
@@ -4294,6 +4295,54 @@ export default {
     },
     builtinTag: "기본제공",
     confirmDelete: '모델 "{name}"을(를) 삭제하시겠습니까?',
+    debug: {
+      title: "모델 테스트",
+      description: "구성된 모델에 실제 요청을 보내고 응답과 소요 시간을 확인합니다",
+      groupModel: "모델 선택",
+      groupInput: "테스트 입력",
+      groupResult: "실행 결과",
+      modelType: "모델 유형",
+      model: "모델",
+      modelPlaceholder: "테스트할 모델을 선택하세요",
+      noModelsForType: "이 유형의 저장된 모델이 없습니다",
+      query: "입력 내용",
+      queryPlaceholder: "모델에 보낼 내용을 입력하세요",
+      embeddingInput: "벡터화할 텍스트",
+      embeddingPlaceholder: "Embedding을 생성할 텍스트를 입력하세요",
+      vlmPrompt: "이미지 프롬프트",
+      vlmPromptPlaceholder: "예: 이 이미지를 자세히 설명해 주세요",
+      documents: "후보 문서",
+      documentsPlaceholder: "후보 문서를 한 줄에 하나씩 입력하세요",
+      documentsHint: "비어 있지 않은 각 줄은 ReRank 모델에 별도 문서로 전송됩니다",
+      imageFile: "이미지 파일",
+      audioFile: "오디오 파일",
+      chooseFile: "파일 선택",
+      parameters: "요청 매개변수",
+      thinking: "사고 모드",
+      thinkingDesc: "사고 모드를 지원하는 모델에만 적용됩니다",
+      systemPrompt: "System Prompt",
+      systemPromptPlaceholder: "선택 사항, 시스템 프롬프트 입력",
+      run: "테스트 실행",
+      copyResult: "결과 복사",
+      history: "기록",
+      thinkOn: "사고 켜짐",
+      thinkOff: "사고 꺼짐",
+      runLabel: "{n}번째 실행",
+      success: "호출 성공",
+      failed: "호출 실패",
+      rawResponse: "응답 내용",
+      requestPreview: "요청 미리보기",
+      requestFailed: "모델 테스트 요청 실패",
+      metrics: {
+        dimension: "벡터 차원",
+        resultCount: "결과 수",
+        answerChars: "답변 문자 수",
+        reasoningChars: "추론 문자 수",
+        reasoningReturned: "추론 내용 반환",
+        textChars: "전사 문자 수",
+        segmentCount: "세그먼트 수",
+      },
+    },
   },
   // Ollama 설정
   ollamaSettings: {
diff --git a/frontend/src/i18n/locales/ru-RU.ts b/frontend/src/i18n/locales/ru-RU.ts
index caa5619d..f0926cf0 100755
--- a/frontend/src/i18n/locales/ru-RU.ts
+++ b/frontend/src/i18n/locales/ru-RU.ts
@@ -3888,6 +3888,7 @@ export default {
     },
     actions: {
       addModel: 'Добавить модель',
+      debugModel: 'Тест модели',
       setDefault: 'Сделать по умолчанию'
     },
     chat: {
@@ -3948,7 +3949,55 @@ export default {
       viewGuide: 'Посмотреть руководство по управлению встроенными моделями'
     },
     builtinTag: 'Встроенная',
-    confirmDelete: 'Удалить модель «{name}»?'
+    confirmDelete: 'Удалить модель «{name}»?',
+    debug: {
+      title: 'Тест модели',
+      description: 'Отправьте реальный запрос к настроенной модели и проверьте ответ и время выполнения',
+      groupModel: 'Выбор модели',
+      groupInput: 'Тестовый ввод',
+      groupResult: 'Результат',
+      modelType: 'Тип модели',
+      model: 'Модель',
+      modelPlaceholder: 'Выберите модель для теста',
+      noModelsForType: 'Нет сохранённых моделей этого типа',
+      query: 'Ввод',
+      queryPlaceholder: 'Введите текст для отправки модели',
+      embeddingInput: 'Текст для векторизации',
+      embeddingPlaceholder: 'Введите текст для создания embedding',
+      vlmPrompt: 'Промпт для изображения',
+      vlmPromptPlaceholder: 'Например: подробно опишите это изображение',
+      documents: 'Кандидаты документов',
+      documentsPlaceholder: 'Введите один документ на строку',
+      documentsHint: 'Каждая непустая строка отправляется в модель ReRank как отдельный документ',
+      imageFile: 'Файл изображения',
+      audioFile: 'Аудиофайл',
+      chooseFile: 'Выбрать файл',
+      parameters: 'Параметры запроса',
+      thinking: 'Режим размышления',
+      thinkingDesc: 'Применяется только к моделям с поддержкой размышления',
+      systemPrompt: 'System Prompt',
+      systemPromptPlaceholder: 'Необязательно: системный промпт',
+      run: 'Запустить тест',
+      copyResult: 'Копировать результат',
+      history: 'История',
+      thinkOn: 'Размышление вкл.',
+      thinkOff: 'Размышление выкл.',
+      runLabel: 'Запуск №{n}',
+      success: 'Запрос выполнен',
+      failed: 'Запрос не выполнен',
+      rawResponse: 'Ответ',
+      requestPreview: 'Просмотр запроса',
+      requestFailed: 'Не удалось выполнить тест модели',
+      metrics: {
+        dimension: 'Размерность вектора',
+        resultCount: 'Количество результатов',
+        answerChars: 'Символов в ответе',
+        reasoningChars: 'Символов рассуждения',
+        reasoningReturned: 'Рассуждение возвращено',
+        textChars: 'Символов транскрипции',
+        segmentCount: 'Количество сегментов',
+      },
+    },
   },
   ollamaSettings: {
     title: 'Настройки Ollama',
diff --git a/frontend/src/i18n/locales/zh-CN.ts b/frontend/src/i18n/locales/zh-CN.ts
index affc406b..2d7fe3d8 100755
--- a/frontend/src/i18n/locales/zh-CN.ts
+++ b/frontend/src/i18n/locales/zh-CN.ts
@@ -4237,6 +4237,7 @@ export default {
     },
     actions: {
       addModel: "添加模型",
+      debugModel: "模型测试",
       setDefault: "设为默认",
     },
     source: {
@@ -4298,6 +4299,54 @@ export default {
     },
     builtinTag: "内置",
     confirmDelete: "确定删除模型「{name}」吗？",
+    debug: {
+      title: "模型测试",
+      description: "向已配置的模型发送真实请求，查看响应与耗时",
+      groupModel: "选择模型",
+      groupInput: "测试输入",
+      groupResult: "运行结果",
+      modelType: "模型类型",
+      model: "模型",
+      modelPlaceholder: "请选择要测试的模型",
+      noModelsForType: "当前类型暂无已添加模型",
+      query: "输入内容",
+      queryPlaceholder: "输入要发送给模型的内容",
+      embeddingInput: "待向量化文本",
+      embeddingPlaceholder: "输入要生成 Embedding 的文本",
+      vlmPrompt: "图片提示词",
+      vlmPromptPlaceholder: "例如：请详细描述这张图片",
+      documents: "候选文档",
+      documentsPlaceholder: "每行输入一个候选文档",
+      documentsHint: "每个非空行会作为一个独立文档发送给 ReRank 模型",
+      imageFile: "图片文件",
+      audioFile: "音频文件",
+      chooseFile: "选择文件",
+      parameters: "请求参数",
+      thinking: "思考模式",
+      thinkingDesc: "仅对支持思考模式的模型生效",
+      systemPrompt: "System Prompt",
+      systemPromptPlaceholder: "可选，输入系统提示词",
+      run: "运行测试",
+      copyResult: "复制结果",
+      history: "历史记录",
+      thinkOn: "思考开启",
+      thinkOff: "思考关闭",
+      runLabel: "第 {n} 次运行",
+      success: "调用成功",
+      failed: "调用失败",
+      rawResponse: "响应内容",
+      requestPreview: "请求预览",
+      requestFailed: "模型测试请求失败",
+      metrics: {
+        dimension: "向量维度",
+        resultCount: "结果数量",
+        answerChars: "回答字符数",
+        reasoningChars: "推理字符数",
+        reasoningReturned: "返回推理内容",
+        textChars: "转写字符数",
+        segmentCount: "分段数量",
+      },
+    },
   },
   // 新增：Ollama 设置
   ollamaSettings: {
diff --git a/frontend/src/utils/request.ts b/frontend/src/utils/request.ts
index b1f3daac..152e9be9 100644
--- a/frontend/src/utils/request.ts
+++ b/frontend/src/utils/request.ts
@@ -257,13 +257,20 @@ export async function getDown(url: string) {
   return res
 }
 
-export function postUpload(url: string, data = {}, onUploadProgress?: (progressEvent: any) => void) {
+export function postUpload(
+  url: string,
+  data = {},
+  onUploadProgress?: (progressEvent: any) => void,
+  config: any = {},
+) {
   return instance.post(url, data, {
+    ...config,
     headers: {
       "Content-Type": "multipart/form-data",
       "X-Request-ID": `${generateRandomString(12)}`,
+      ...(config.headers || {}),
     },
-    onUploadProgress,
+    onUploadProgress: onUploadProgress || config.onUploadProgress,
   });
 }
 
diff --git a/frontend/src/utils/thinkingControl.ts b/frontend/src/utils/thinkingControl.ts
index 97723617..122bced7 100644
--- a/frontend/src/utils/thinkingControl.ts
+++ b/frontend/src/utils/thinkingControl.ts
@@ -69,3 +69,21 @@ export function resolveThinkingControl(
   }
   return defaultThinkingControl(provider, modelName)
 }
+
+/** Whether the debug drawer should expose Think on/off for this saved model. */
+export function modelSupportsThinking(model: {
+  type: string
+  source: string
+  name: string
+  parameters: {
+    provider?: string
+    extra_config?: { thinking_control?: string }
+  }
+}): boolean {
+  if (model.type !== 'KnowledgeQA' || model.source !== 'remote') return false
+  return resolveThinkingControl(
+    model.parameters.extra_config?.thinking_control,
+    model.parameters.provider || '',
+    model.name || '',
+  ) !== 'none'
+}
diff --git a/frontend/src/views/settings/ModelSettings.vue b/frontend/src/views/settings/ModelSettings.vue
index e6a1342d..0994a0db 100644
--- a/frontend/src/views/settings/ModelSettings.vue
+++ b/frontend/src/views/settings/ModelSettings.vue
@@ -1,8 +1,24 @@
 <template>
   <div class="model-settings">
     <div class="section-header">
-      <h2>{{ $t('modelSettings.title') }}</h2>
-      <p class="section-description">{{ $t('modelSettings.description') }}</p>
+      <div class="section-header__top">
+        <div>
+          <h2>{{ $t('modelSettings.title') }}</h2>
+          <p class="section-description">{{ $t('modelSettings.description') }}</p>
+        </div>
+        <t-button
+          v-if="authStore.hasRole('admin')"
+          type="button"
+          theme="primary"
+          variant="text"
+          size="medium"
+          class="model-test-trigger"
+          @click="showDebugDrawer = true"
+        >
+          <template #icon><play-circle-icon /></template>
+          {{ $t('modelSettings.actions.debugModel') }}
+        </t-button>
+      </div>
 
       <div class="builtin-models-hint" role="note">
         <p class="builtin-hint-label">{{ $t('modelSettings.builtinModels.title') }}</p>
@@ -114,6 +130,7 @@
     <!-- 模型编辑器抽屉 -->
     <ModelEditorDialog v-model:visible="showDialog" :model-type="currentModelType" :model-data="editingModel"
       @confirm="handleModelSave" />
+    <ModelDebugDrawer v-model:visible="showDebugDrawer" :models="allModels" />
 
   </div>
 </template>
@@ -121,9 +138,10 @@
 <script setup lang="ts">
 import { ref, computed, onMounted } from 'vue'
 import { MessagePlugin } from 'tdesign-vue-next'
-import { AddIcon } from 'tdesign-icons-vue-next'
+import { AddIcon, PlayCircleIcon } from 'tdesign-icons-vue-next'
 import { useI18n } from 'vue-i18n'
 import ModelEditorDialog from '@/components/ModelEditorDialog.vue'
+import ModelDebugDrawer from '@/components/ModelDebugDrawer.vue'
 import { listModels, createModel, updateModel as updateModelAPI, deleteModel as deleteModelAPI, type ModelConfig } from '@/api/model'
 import { useAuthStore } from '@/stores/auth'
 
@@ -133,6 +151,7 @@ type ModelType = 'chat' | 'embedding' | 'rerank' | 'vllm' | 'asr'
 type FilterType = 'all' | ModelType
 
 const showDialog = ref(false)
+const showDebugDrawer = ref(false)
 const currentModelType = ref<ModelType>('chat')
 const editingModel = ref<any>(null)
 const loading = ref(true)
@@ -585,6 +604,33 @@ onMounted(() => {
   }
 }
 
+.section-header__top {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 20px;
+}
+
+.model-test-trigger {
+  --td-bg-color-container-hover: transparent;
+  flex-shrink: 0;
+  padding-left: 0;
+  padding-right: 0;
+  font-weight: 600;
+
+  &:hover,
+  &:focus,
+  &.t-is-active,
+  &:active {
+    background-color: transparent !important;
+    color: var(--td-brand-color-hover);
+  }
+
+  &:active {
+    color: var(--td-brand-color-active);
+  }
+}
+
 .builtin-models-hint {
   margin-top: 12px;
   padding: 10px 12px;
diff --git a/internal/handler/model.go b/internal/handler/model.go
index 0292aefa..fe73492a 100644
--- a/internal/handler/model.go
+++ b/internal/handler/model.go
@@ -1,12 +1,18 @@
 package handler
 
 import (
+	"encoding/json"
+	"fmt"
+	"io"
 	"net/http"
+	"strings"
+	"time"
 
 	"github.com/Tencent/WeKnora/internal/application/service"
 	"github.com/Tencent/WeKnora/internal/errors"
 	"github.com/Tencent/WeKnora/internal/handler/dto"
 	"github.com/Tencent/WeKnora/internal/logger"
+	"github.com/Tencent/WeKnora/internal/models/chat"
 	"github.com/Tencent/WeKnora/internal/models/provider"
 	"github.com/Tencent/WeKnora/internal/types"
 	"github.com/Tencent/WeKnora/internal/types/interfaces"
@@ -199,6 +205,323 @@ func (h *ModelHandler) ListModels(c *gin.Context) {
 	})
 }
 
+const (
+	modelDebugMaxInputBytes = 64 * 1024
+	modelDebugMaxFileBytes  = 20 * 1024 * 1024
+)
+
+// ModelDebugOptions contains the cross-provider parameters exposed by the
+// model debugger. Pointer fields preserve explicit zero/false values.
+type ModelDebugOptions struct {
+	SystemPrompt string   `json:"system_prompt,omitempty"`
+	Temperature  *float64 `json:"temperature,omitempty"`
+	TopP         *float64 `json:"top_p,omitempty"`
+	MaxTokens    *int     `json:"max_tokens,omitempty"`
+	Thinking     *bool    `json:"thinking,omitempty"`
+}
+
+func parseModelDebugOptions(raw string) (ModelDebugOptions, error) {
+	var opts ModelDebugOptions
+	if strings.TrimSpace(raw) == "" {
+		return opts, nil
+	}
+	if err := json.Unmarshal([]byte(raw), &opts); err != nil {
+		return opts, fmt.Errorf("invalid options: %w", err)
+	}
+	if opts.MaxTokens != nil && (*opts.MaxTokens < 1 || *opts.MaxTokens > 8192) {
+		return opts, fmt.Errorf("max_tokens must be between 1 and 8192")
+	}
+	if opts.Temperature != nil && (*opts.Temperature < 0 || *opts.Temperature > 2) {
+		return opts, fmt.Errorf("temperature must be between 0 and 2")
+	}
+	if opts.TopP != nil && (*opts.TopP <= 0 || *opts.TopP > 1) {
+		return opts, fmt.Errorf("top_p must be greater than 0 and at most 1")
+	}
+	return opts, nil
+}
+
+func redactedDebugConfig(config map[string]string) map[string]string {
+	if len(config) == 0 {
+		return nil
+	}
+	out := make(map[string]string, len(config))
+	for key, value := range config {
+		lower := strings.ToLower(key)
+		if strings.Contains(lower, "secret") ||
+			strings.Contains(lower, "token") ||
+			strings.Contains(lower, "password") ||
+			strings.Contains(lower, "api_key") ||
+			strings.Contains(lower, "apikey") ||
+			strings.Contains(lower, "authorization") {
+			out[key] = "[REDACTED]"
+			continue
+		}
+		out[key] = value
+	}
+	return out
+}
+
+func modelDebugRequestPreview(model *types.Model, input string, documents []string, opts ModelDebugOptions, fileName string, fileSize int64) gin.H {
+	preview := gin.H{
+		"model_id":   model.ID,
+		"model_name": model.Name,
+		"model_type": model.Type,
+		"source":     model.Source,
+		"provider":   model.Parameters.Provider,
+		"input":      input,
+		"options":    opts,
+	}
+	if len(documents) > 0 {
+		preview["documents"] = documents
+	}
+	if fileName != "" {
+		preview["file"] = gin.H{"name": fileName, "size": fileSize}
+	}
+	if model.Parameters.ExtraConfig != nil {
+		preview["model_extra_config"] = redactedDebugConfig(model.Parameters.ExtraConfig)
+	}
+	if len(model.Parameters.CustomHeaders) > 0 {
+		headerNames := make([]string, 0, len(model.Parameters.CustomHeaders))
+		for name := range model.Parameters.CustomHeaders {
+			headerNames = append(headerNames, name)
+		}
+		preview["custom_header_names"] = headerNames
+	}
+	return preview
+}
+
+func writeModelDebugResult(c *gin.Context, started time.Time, request gin.H, response any, callErr error, observations gin.H) {
+	data := gin.H{
+		"ok":           callErr == nil,
+		"elapsed_ms":   time.Since(started).Milliseconds(),
+		"request":      request,
+		"raw_response": response,
+		"observations": observations,
+	}
+	if callErr != nil {
+		data["error"] = callErr.Error()
+	}
+	c.JSON(http.StatusOK, gin.H{"success": true, "data": data})
+}
+
+type modelDebugChatStreamResponse struct {
+	Content          string                 `json:"content"`
+	ReasoningContent string                 `json:"reasoning_content,omitempty"`
+	ToolCalls        []types.LLMToolCall    `json:"tool_calls,omitempty"`
+	FinishReason     string                 `json:"finish_reason,omitempty"`
+	Usage            *types.TokenUsage      `json:"usage,omitempty"`
+	StreamEvents     []types.StreamResponse `json:"stream_events"`
+}
+
+func consumeModelDebugChatStream(stream <-chan types.StreamResponse) (*modelDebugChatStreamResponse, error) {
+	result := &modelDebugChatStreamResponse{
+		StreamEvents: make([]types.StreamResponse, 0),
+	}
+	for event := range stream {
+		result.StreamEvents = append(result.StreamEvents, event)
+		switch event.ResponseType {
+		case types.ResponseTypeThinking:
+			result.ReasoningContent += event.Content
+		case types.ResponseTypeAnswer:
+			result.Content += event.Content
+			if len(event.ToolCalls) > 0 {
+				result.ToolCalls = event.ToolCalls
+			}
+			if event.FinishReason != "" {
+				result.FinishReason = event.FinishReason
+			}
+			if event.Usage != nil {
+				result.Usage = event.Usage
+			}
+		case types.ResponseTypeToolCall:
+			if len(event.ToolCalls) > 0 {
+				result.ToolCalls = event.ToolCalls
+			}
+		case types.ResponseTypeError:
+			return result, fmt.Errorf("%s", event.Content)
+		}
+	}
+	return result, nil
+}
+
+// DebugModel executes a saved model through the same service constructors used
+// by production calls and returns the complete normalized response. Credentials
+// stay server-side; the request preview contains only non-secret fields.
+func (h *ModelHandler) DebugModel(c *gin.Context) {
+	ctx := c.Request.Context()
+	started := time.Now()
+	id := secutils.SanitizeForLog(c.Param("id"))
+	if id == "" {
+		c.Error(errors.NewBadRequestError("Model ID cannot be empty"))
+		return
+	}
+
+	model, err := h.service.GetModelByID(ctx, id)
+	if err != nil {
+		if err == service.ErrModelNotFound {
+			c.Error(errors.NewNotFoundError("Model not found"))
+			return
+		}
+		c.Error(errors.NewInternalServerError(err.Error()))
+		return
+	}
+
+	input := c.PostForm("input")
+	if len(input) > modelDebugMaxInputBytes {
+		c.Error(errors.NewBadRequestError("input is too long"))
+		return
+	}
+	opts, err := parseModelDebugOptions(c.PostForm("options"))
+	if err != nil {
+		c.Error(errors.NewBadRequestError(err.Error()))
+		return
+	}
+	var documents []string
+	if rawDocuments := c.PostForm("documents"); strings.TrimSpace(rawDocuments) != "" {
+		if err := json.Unmarshal([]byte(rawDocuments), &documents); err != nil {
+			c.Error(errors.NewBadRequestError("documents must be a JSON string array"))
+			return
+		}
+		if len(documents) > 100 {
+			c.Error(errors.NewBadRequestError("documents cannot exceed 100 items"))
+			return
+		}
+	}
+
+	var (
+		fileBytes []byte
+		fileName  string
+		fileSize  int64
+	)
+	if file, header, fileErr := c.Request.FormFile("file"); fileErr == nil {
+		defer file.Close()
+		fileName = header.Filename
+		fileSize = header.Size
+		if fileSize > modelDebugMaxFileBytes {
+			c.Error(errors.NewBadRequestError("file cannot exceed 20 MB"))
+			return
+		}
+		fileBytes, err = io.ReadAll(io.LimitReader(file, modelDebugMaxFileBytes+1))
+		if err != nil {
+			c.Error(errors.NewBadRequestError("failed to read uploaded file"))
+			return
+		}
+		if len(fileBytes) > modelDebugMaxFileBytes {
+			c.Error(errors.NewBadRequestError("file cannot exceed 20 MB"))
+			return
+		}
+		fileSize = int64(len(fileBytes))
+	}
+
+	requestPreview := modelDebugRequestPreview(model, input, documents, opts, fileName, fileSize)
+	observations := gin.H{}
+
+	switch model.Type {
+	case types.ModelTypeKnowledgeQA:
+		if strings.TrimSpace(input) == "" {
+			c.Error(errors.NewBadRequestError("query cannot be empty"))
+			return
+		}
+		instance, callErr := h.service.GetChatModel(ctx, id)
+		if callErr != nil {
+			writeModelDebugResult(c, started, requestPreview, nil, callErr, observations)
+			return
+		}
+		messages := make([]chat.Message, 0, 2)
+		if strings.TrimSpace(opts.SystemPrompt) != "" {
+			messages = append(messages, chat.Message{Role: "system", Content: opts.SystemPrompt})
+		}
+		messages = append(messages, chat.Message{Role: "user", Content: input})
+		chatOpts := &chat.ChatOptions{}
+		if opts.Temperature != nil {
+			chatOpts.Temperature = *opts.Temperature
+		}
+		if opts.TopP != nil {
+			chatOpts.TopP = *opts.TopP
+		}
+		if opts.MaxTokens != nil {
+			chatOpts.MaxTokens = *opts.MaxTokens
+		}
+		chatOpts.Thinking = opts.Thinking
+		chatConfig := chat.ConfigFromModel(model, "", "")
+		thinkingControl := chat.EffectiveThinkingControl(chatConfig)
+		observations["stream"] = true
+		observations["requested_thinking"] = opts.Thinking != nil && *opts.Thinking
+		observations["thinking_control"] = thinkingControl
+		observations["thinking_parameter_sent"] = opts.Thinking != nil && thinkingControl != "none"
+
+		stream, callErr := instance.ChatStream(ctx, messages, chatOpts)
+		if callErr != nil {
+			writeModelDebugResult(c, started, requestPreview, nil, callErr, observations)
+			return
+		}
+		resp, callErr := consumeModelDebugChatStream(stream)
+		if resp != nil {
+			observations["reasoning_returned"] = strings.TrimSpace(resp.ReasoningContent) != ""
+			observations["reasoning_characters"] = len([]rune(resp.ReasoningContent))
+			observations["answer_characters"] = len([]rune(resp.Content))
+		}
+		writeModelDebugResult(c, started, requestPreview, resp, callErr, observations)
+	case types.ModelTypeEmbedding:
+		if strings.TrimSpace(input) == "" {
+			c.Error(errors.NewBadRequestError("input cannot be empty"))
+			return
+		}
+		instance, callErr := h.service.GetEmbeddingModel(ctx, id)
+		if callErr != nil {
+			writeModelDebugResult(c, started, requestPreview, nil, callErr, observations)
+			return
+		}
+		vector, callErr := instance.Embed(ctx, input)
+		observations["dimension"] = len(vector)
+		writeModelDebugResult(c, started, requestPreview, vector, callErr, observations)
+	case types.ModelTypeRerank:
+		if strings.TrimSpace(input) == "" || len(documents) == 0 {
+			c.Error(errors.NewBadRequestError("query and documents cannot be empty"))
+			return
+		}
+		instance, callErr := h.service.GetRerankModel(ctx, id)
+		if callErr != nil {
+			writeModelDebugResult(c, started, requestPreview, nil, callErr, observations)
+			return
+		}
+		results, callErr := instance.Rerank(ctx, input, documents)
+		observations["result_count"] = len(results)
+		writeModelDebugResult(c, started, requestPreview, results, callErr, observations)
+	case types.ModelTypeVLLM:
+		if len(fileBytes) == 0 {
+			c.Error(errors.NewBadRequestError("image file is required"))
+			return
+		}
+		instance, callErr := h.service.GetVLMModel(ctx, id)
+		if callErr != nil {
+			writeModelDebugResult(c, started, requestPreview, nil, callErr, observations)
+			return
+		}
+		result, callErr := instance.Predict(ctx, [][]byte{fileBytes}, input)
+		observations["answer_characters"] = len([]rune(result))
+		writeModelDebugResult(c, started, requestPreview, result, callErr, observations)
+	case types.ModelTypeASR:
+		if len(fileBytes) == 0 {
+			c.Error(errors.NewBadRequestError("audio file is required"))
+			return
+		}
+		instance, callErr := h.service.GetASRModel(ctx, id)
+		if callErr != nil {
+			writeModelDebugResult(c, started, requestPreview, nil, callErr, observations)
+			return
+		}
+		result, callErr := instance.Transcribe(ctx, fileBytes, fileName)
+		if result != nil {
+			observations["text_characters"] = len([]rune(result.Text))
+			observations["segment_count"] = len(result.Segments)
+		}
+		writeModelDebugResult(c, started, requestPreview, result, callErr, observations)
+	default:
+		c.Error(errors.NewBadRequestError("unsupported model type"))
+	}
+}
+
 // UpdateModelRequest defines the structure for model update requests
 // Contains fields that can be updated for an existing model
 type UpdateModelRequest struct {
diff --git a/internal/handler/model_test.go b/internal/handler/model_test.go
index 7abed79a..298b8ece 100644
--- a/internal/handler/model_test.go
+++ b/internal/handler/model_test.go
@@ -4,6 +4,7 @@ import (
 	"encoding/json"
 	"testing"
 
+	"github.com/Tencent/WeKnora/internal/types"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -18,3 +19,54 @@ func TestModelUpdateRequestDisplayNamePresence(t *testing.T) {
 	require.NotNil(t, cleared.DisplayName)
 	assert.Equal(t, "", *cleared.DisplayName)
 }
+
+func TestParseModelDebugOptionsPreservesExplicitThinkingFalse(t *testing.T) {
+	opts, err := parseModelDebugOptions(`{"thinking":false,"temperature":0,"max_tokens":256}`)
+	require.NoError(t, err)
+	require.NotNil(t, opts.Thinking)
+	assert.False(t, *opts.Thinking)
+	require.NotNil(t, opts.Temperature)
+	assert.Zero(t, *opts.Temperature)
+	require.NotNil(t, opts.MaxTokens)
+	assert.Equal(t, 256, *opts.MaxTokens)
+}
+
+func TestParseModelDebugOptionsRejectsOutOfRangeValues(t *testing.T) {
+	_, err := parseModelDebugOptions(`{"top_p":0}`)
+	require.ErrorContains(t, err, "top_p")
+}
+
+func TestRedactedDebugConfig(t *testing.T) {
+	got := redactedDebugConfig(map[string]string{
+		"thinking_control": "enable_thinking",
+		"secret_key":       "do-not-leak",
+		"access_token":     "do-not-leak-either",
+	})
+	assert.Equal(t, "enable_thinking", got["thinking_control"])
+	assert.Equal(t, "[REDACTED]", got["secret_key"])
+	assert.Equal(t, "[REDACTED]", got["access_token"])
+}
+
+func TestConsumeModelDebugChatStream(t *testing.T) {
+	stream := make(chan types.StreamResponse, 5)
+	stream <- types.StreamResponse{ResponseType: types.ResponseTypeThinking, Content: "reason "}
+	stream <- types.StreamResponse{ResponseType: types.ResponseTypeThinking, Content: "more", Done: true}
+	stream <- types.StreamResponse{ResponseType: types.ResponseTypeAnswer, Content: "answer "}
+	stream <- types.StreamResponse{ResponseType: types.ResponseTypeAnswer, Content: "done"}
+	stream <- types.StreamResponse{
+		ResponseType: types.ResponseTypeAnswer,
+		Done:         true,
+		FinishReason: "stop",
+		Usage:        &types.TokenUsage{PromptTokens: 3, CompletionTokens: 4, TotalTokens: 7},
+	}
+	close(stream)
+
+	got, err := consumeModelDebugChatStream(stream)
+	require.NoError(t, err)
+	assert.Equal(t, "reason more", got.ReasoningContent)
+	assert.Equal(t, "answer done", got.Content)
+	assert.Equal(t, "stop", got.FinishReason)
+	require.NotNil(t, got.Usage)
+	assert.Equal(t, 7, got.Usage.TotalTokens)
+	assert.Len(t, got.StreamEvents, 5)
+}
diff --git a/internal/models/chat/thinking.go b/internal/models/chat/thinking.go
index f676208d..23e92372 100644
--- a/internal/models/chat/thinking.go
+++ b/internal/models/chat/thinking.go
@@ -3,6 +3,7 @@ package chat
 import (
 	"strings"
 
+	"github.com/Tencent/WeKnora/internal/models/provider"
 	"github.com/sashabaranov/go-openai"
 )
 
@@ -138,3 +139,34 @@ func parseThinkingOverride(extraConfig map[string]string) ThinkingStrategy {
 		return chatTemplateKwargs{}
 	}
 }
+
+// EffectiveThinkingControl reports the provider field that will carry
+// ChatOptions.Thinking. It intentionally shares the same adapter/override
+// resolution as the real request path so diagnostics do not guess from the
+// frontend selection.
+func EffectiveThinkingControl(config *ChatConfig) string {
+	if config == nil {
+		return "none"
+	}
+	if override := parseThinkingOverride(config.ExtraConfig); override != nil {
+		return thinkingStrategyName(override)
+	}
+	providerName := provider.ProviderName(config.Provider)
+	if providerName == "" {
+		providerName = provider.DetectProvider(config.BaseURL)
+	}
+	return thinkingStrategyName(resolveProvider(providerName, config.ModelName).Thinking())
+}
+
+func thinkingStrategyName(strategy ThinkingStrategy) string {
+	switch strategy.(type) {
+	case enableThinking:
+		return "enable_thinking"
+	case thinkingTypeField:
+		return "thinking_type"
+	case chatTemplateKwargs:
+		return "chat_template_kwargs"
+	default:
+		return "none"
+	}
+}
diff --git a/internal/models/chat/thinking_test.go b/internal/models/chat/thinking_test.go
index d55ad0a1..13996997 100644
--- a/internal/models/chat/thinking_test.go
+++ b/internal/models/chat/thinking_test.go
@@ -125,3 +125,20 @@ func TestParseThinkingOverride(t *testing.T) {
 	assert.Nil(t, parseThinkingOverride(map[string]string{}))
 	assert.Nil(t, parseThinkingOverride(map[string]string{ExtraConfigThinkingControl: ""}))
 }
+
+func TestEffectiveThinkingControl(t *testing.T) {
+	assert.Equal(t, "enable_thinking", EffectiveThinkingControl(&ChatConfig{
+		Provider:  "aliyun",
+		ModelName: "qwen3-32b",
+	}))
+	assert.Equal(t, "chat_template_kwargs", EffectiveThinkingControl(&ChatConfig{
+		Provider:    "generic",
+		ModelName:   "qwen3",
+		ExtraConfig: map[string]string{ExtraConfigThinkingControl: "chat_template_kwargs"},
+	}))
+	assert.Equal(t, "none", EffectiveThinkingControl(&ChatConfig{
+		Provider:    "generic",
+		ModelName:   "qwen3",
+		ExtraConfig: map[string]string{ExtraConfigThinkingControl: "none"},
+	}))
+}
diff --git a/internal/router/router.go b/internal/router/router.go
index 2d2eb8ba..5dca8911 100644
--- a/internal/router/router.go
+++ b/internal/router/router.go
@@ -635,6 +635,8 @@ func RegisterModelRoutes(
 		models.POST("", g.Admin(), handler.CreateModel)
 		// 获取模型列表 — Viewer+
 		models.GET("", g.Viewer(), handler.ListModels)
+		// 调试已保存模型会发起真实上游调用并产生费用 — Admin+
+		models.POST("/:id/debug", g.Admin(), handler.DebugModel)
 		// 获取单个模型 — Viewer+
 		models.GET("/:id", g.Viewer(), handler.GetModel)
 		// 更新模型 — Admin+