diff --git a/.sync-watermark b/.sync-watermark
index 87c20b5..d8156dc 100644
--- a/.sync-watermark
+++ b/.sync-watermark
@@ -4,6 +4,6 @@
 # 下次增量同步时，只处理 OpenRouter 中 created > 此值的新模型。
 # 
 # 格式: Unix timestamp（秒）
-# 人读: 2026-06-24T04:45:03Z
+# 人读: 2026-07-21T03:37:38Z
 # 
-1782276303
+1784554658
diff --git a/__tests__/validate.test.mjs b/__tests__/validate.test.mjs
index 53ae7ce..3469373 100644
--- a/__tests__/validate.test.mjs
+++ b/__tests__/validate.test.mjs
@@ -271,8 +271,101 @@ describe('真实数据全量校验', () => {
 
     assert.equal(specs.length, 1, 'mimo-v2.5-asr 应且仅应有一条规格')
     assert.deepEqual(specs[0].match.exact, ['mimo-v2.5-asr'])
+    assert.equal(specs[0].spec.contextWindow, 8192)
+    assert.equal(specs[0].spec.maxOutputTokens, 2048)
     assert.deepEqual(specs[0].spec.serviceType, ['asr'])
     assert.ok(specs[0].spec.capabilities.includes('asr'))
+    assert.deepEqual(specs[0].spec.extra.modelOrigin, {
+      vendor: 'xiaomi',
+      canonicalModelId: 'mimo-v2.5-asr',
+      family: 'mimo-asr',
+    })
+    assert.deepEqual(specs[0].spec.extra.speech, {
+      protocol: 'xiaomi-mimo-chat-asr',
+      protocolVersion: '1',
+      modes: ['batch'],
+      inputFormats: ['wav', 'mp3'],
+      languageHints: true,
+    })
+  })
+
+  it('MiMo V2.5 TTS 应声明稳定协议与官方音色', () => {
+    const specFile = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'xiaomi.json'), 'utf8'))
+    const provider = JSON.parse(readFileSync(join(ROOT, 'compute', 'providers', 'xiaomi.json'), 'utf8'))
+    const tts = specFile.specs.find((item) => item.id === 'mimo-v2.5-tts')
+    const providerTts = provider.models.find((item) => item.modelName === 'mimo-v2.5-tts')
+    const providerAsr = provider.models.find((item) => item.modelName === 'mimo-v2.5-asr')
+
+    assert.ok(tts, 'model-specs 中缺少 mimo-v2.5-tts')
+    assert.ok(providerTts, 'provider 中缺少 mimo-v2.5-tts')
+    assert.ok(providerAsr, 'provider 中缺少 mimo-v2.5-asr')
+    assert.ok(provider.services.includes('asr'))
+    assert.equal(tts.spec.contextWindow, 8192)
+    assert.equal(tts.spec.maxOutputTokens, 8192)
+    assert.deepEqual(tts.spec.extra.modelOrigin, {
+      vendor: 'xiaomi',
+      canonicalModelId: 'mimo-v2.5-tts',
+      family: 'mimo-tts',
+    })
+    assert.equal(tts.spec.extra.speech.protocol, 'xiaomi-mimo-chat-tts')
+    assert.equal(tts.spec.extra.speech.protocolVersion, '1')
+    assert.deepEqual(tts.spec.extra.speech.modes, ['batch', 'streaming'])
+    assert.deepEqual(tts.spec.extra.speech.outputFormats, ['wav', 'mp3', 'pcm16le'])
+    assert.deepEqual(tts.spec.extra.speech.sampleRates, [24000])
+    assert.equal(tts.spec.extra.speech.defaultVoice, 'mimo_default')
+    assert.deepEqual(
+      tts.spec.extra.speech.voices.map((voice) => voice.id),
+      ['mimo_default', '冰糖', '茉莉', '苏打', '白桦', 'Mia', 'Chloe', 'Milo', 'Dean'],
+    )
+    assert.deepEqual(
+      providerTts.extra.voices.map((voice) => voice.id),
+      ['mimo_default', '冰糖', '茉莉', '苏打', '白桦', 'Mia', 'Chloe', 'Milo', 'Dean'],
+    )
+    assert.deepEqual(providerTts.extra.audioFormats, ['wav', 'mp3', 'pcm16le'])
+    assert.equal(providerTts.extra.sampleRate, 24000)
+  })
+
+  it('OpenRouter 2026-07-21 水位线前的缺失模型应使用当前精确规格', () => {
+    const openai = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'openai.json'), 'utf8'))
+    const google = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'google.json'), 'utf8'))
+    const tencent = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'tencent.json'), 'utf8'))
+    const xai = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'xai.json'), 'utf8'))
+
+    for (const id of ['gpt-5.6-sol-pro', 'gpt-5.6-terra-pro', 'gpt-5.6-luna-pro']) {
+      const modelSpec = openai.specs.find((item) => item.id === id)
+      assert.ok(modelSpec, `model-specs/openai.json 缺少 ${id}`)
+      assert.equal(modelSpec.family, id)
+      assert.deepEqual(modelSpec.match.exact, [id])
+      assert.equal('patterns' in modelSpec.match, false)
+      assert.equal(modelSpec.spec.contextWindow, 1050000)
+      assert.equal(modelSpec.spec.maxOutputTokens, 128000)
+      assert.equal(modelSpec.spec.supportsReasoning, true)
+    }
+
+    const gemini = google.specs.find((item) => item.id === 'gemini-3.1-flash-lite-image')
+    assert.ok(gemini, 'model-specs/google.json 缺少 gemini-3.1-flash-lite-image')
+    assert.deepEqual(gemini.match.exact, ['gemini-3.1-flash-lite-image'])
+    assert.equal(gemini.spec.contextWindow, 65536)
+    assert.equal(gemini.spec.maxOutputTokens, 4096)
+    assert.ok(gemini.spec.capabilities.includes('image_generation'))
+    assert.ok(gemini.spec.capabilities.includes('image_editing'))
+
+    const hy3 = tencent.specs.find((item) => item.id === 'hy3')
+    assert.ok(hy3, 'model-specs/tencent.json 缺少 hy3')
+    assert.deepEqual(hy3.match.exact, ['hy3'])
+    assert.equal(hy3.spec.contextWindow, 262144)
+    assert.equal(hy3.spec.maxOutputTokens, 128000)
+
+    const grok = xai.specs.find((item) => item.id === 'grok-4.5')
+    assert.ok(grok, 'model-specs/xai.json 缺少 grok-4.5')
+    assert.deepEqual(grok.match.exact, ['grok-4.5'])
+    assert.equal(grok.spec.contextWindow, 500000)
+    assert.equal(grok.spec.maxOutputTokens, 450000)
+    assert.equal(grok.spec.extra.thinkingOnly, true)
+
+    const watermark = readFileSync(join(ROOT, '.sync-watermark'), 'utf8')
+    assert.match(watermark, /# 人读: 2026-07-21T03:37:38Z/)
+    assert.equal(Number(watermark.trim().split(/\s+/).at(-1)), 1784554658)
   })
 
   it('Ox Alpha 应提供与 OpenRouter 一致的多模态推理规格', () => {
@@ -751,6 +844,64 @@ describe('model-spec schema 接入面边界', () => {
     }
     assert.equal(validate(data), true, JSON.stringify(validate.errors))
   })
+
+  it('接受 extra.modelOrigin/speech，同时保持其他扩展开放', () => {
+    const data = {
+      specs: [{
+        id: 'mimo-v2.5-asr',
+        spec: {
+          extra: {
+            modelOrigin: {
+              vendor: 'xiaomi',
+              canonicalModelId: 'mimo-v2.5-asr',
+              family: 'mimo-asr',
+            },
+            speech: {
+              protocol: 'xiaomi-mimo-chat-asr',
+              protocolVersion: '1',
+              modes: ['batch', 'streaming'],
+              inputFormats: ['wav', 'mp3'],
+              languageHints: true,
+            },
+            intrinsicBudgetHint: 8192,
+          },
+        },
+      }],
+    }
+    assert.equal(validate(data), true, JSON.stringify(validate.errors))
+  })
+
+  it('严格拒绝非法 speech 二级结构和字段越界', () => {
+    const invalidProfile = {
+      specs: [{
+        id: 'bad-asr',
+        spec: {
+          extra: {
+            speech: {
+              protocol: 'xiaomi-mimo-chat-asr',
+              protocolVersion: '1',
+              modes: ['batch', 'batch'],
+            },
+          },
+        },
+      }],
+    }
+    assert.equal(validate(invalidProfile), false)
+
+    const misplaced = {
+      specs: [{
+        id: 'bad-asr',
+        spec: {
+          speech: {
+            protocol: 'xiaomi-mimo-chat-asr',
+            protocolVersion: '1',
+            modes: ['batch'],
+          },
+        },
+      }],
+    }
+    assert.equal(validate(misplaced), false)
+  })
 })
 
 // ==================== Manifest schema ====================
diff --git a/compute/model-specs/google.json b/compute/model-specs/google.json
index 651eaf2..54e3b40 100644
--- a/compute/model-specs/google.json
+++ b/compute/model-specs/google.json
@@ -1,6 +1,33 @@
 {
   "description": "Google Gemini 系列模型规格。参数来源：config-center compute/providers/google.json。",
   "specs": [
+    {
+      "id": "gemini-3.1-flash-lite-image",
+      "displayName": "Gemini 3.1 Flash Lite Image (Nano Banana 2 Lite)",
+      "family": "gemini-3.1-flash-lite-image",
+      "match": {
+        "exact": [
+          "gemini-3.1-flash-lite-image"
+        ]
+      },
+      "spec": {
+        "contextWindow": 65536,
+        "maxOutputTokens": 4096,
+        "capabilities": [
+          "chat",
+          "vision",
+          "image_generation",
+          "image_editing",
+          "tool_use"
+        ],
+        "serviceType": [
+          "image_gen"
+        ],
+        "supportsReasoning": true,
+        "description": "Nano Banana 2 Lite，Google 面向高吞吐图像生成与编辑的轻量模型",
+        "releasedAt": "2026-06-30"
+      }
+    },
     {
       "id": "gemini-3.5-flash",
       "displayName": "Gemini 3.5 Flash",
diff --git a/compute/model-specs/openai.json b/compute/model-specs/openai.json
index 1cbf39e..a101e76 100644
--- a/compute/model-specs/openai.json
+++ b/compute/model-specs/openai.json
@@ -749,6 +749,35 @@
         }
       }
     },
+    {
+      "id": "gpt-5.6-sol-pro",
+      "displayName": "GPT-5.6 Sol Pro",
+      "family": "gpt-5.6-sol-pro",
+      "match": {
+        "exact": [
+          "gpt-5.6-sol-pro"
+        ]
+      },
+      "spec": {
+        "contextWindow": 1050000,
+        "maxOutputTokens": 128000,
+        "capabilities": [
+          "chat",
+          "reasoning",
+          "code",
+          "vision",
+          "long_context",
+          "tool_use",
+          "agent"
+        ],
+        "serviceType": [
+          "chat"
+        ],
+        "supportsReasoning": true,
+        "description": "GPT-5.6 Sol 的 Pro 固定推理变体，由上游以 reasoning.mode=pro 提供",
+        "releasedAt": "2026-07-09"
+      }
+    },
     {
       "id": "gpt-5.6-terra",
       "displayName": "GPT-5.6 Terra",
@@ -788,6 +817,35 @@
         }
       }
     },
+    {
+      "id": "gpt-5.6-terra-pro",
+      "displayName": "GPT-5.6 Terra Pro",
+      "family": "gpt-5.6-terra-pro",
+      "match": {
+        "exact": [
+          "gpt-5.6-terra-pro"
+        ]
+      },
+      "spec": {
+        "contextWindow": 1050000,
+        "maxOutputTokens": 128000,
+        "capabilities": [
+          "chat",
+          "reasoning",
+          "code",
+          "vision",
+          "long_context",
+          "tool_use",
+          "agent"
+        ],
+        "serviceType": [
+          "chat"
+        ],
+        "supportsReasoning": true,
+        "description": "GPT-5.6 Terra 的 Pro 固定推理变体，由上游以 reasoning.mode=pro 提供",
+        "releasedAt": "2026-07-09"
+      }
+    },
     {
       "id": "gpt-5.6-luna",
       "displayName": "GPT-5.6 Luna",
@@ -827,6 +885,36 @@
         }
       }
     },
+    {
+      "id": "gpt-5.6-luna-pro",
+      "displayName": "GPT-5.6 Luna Pro",
+      "family": "gpt-5.6-luna-pro",
+      "match": {
+        "exact": [
+          "gpt-5.6-luna-pro"
+        ]
+      },
+      "spec": {
+        "contextWindow": 1050000,
+        "maxOutputTokens": 128000,
+        "capabilities": [
+          "chat",
+          "reasoning",
+          "code",
+          "vision",
+          "long_context",
+          "tool_use",
+          "agent",
+          "fast"
+        ],
+        "serviceType": [
+          "chat"
+        ],
+        "supportsReasoning": true,
+        "description": "GPT-5.6 Luna 的 Pro 固定推理变体，由上游以 reasoning.mode=pro 提供",
+        "releasedAt": "2026-07-09"
+      }
+    },
     {
       "id": "gpt-5.4-mini",
       "displayName": "GPT-5.4 Mini",
diff --git a/compute/model-specs/tencent.json b/compute/model-specs/tencent.json
index bdd75f2..8a3bec9 100644
--- a/compute/model-specs/tencent.json
+++ b/compute/model-specs/tencent.json
@@ -1,6 +1,35 @@
 {
   "description": "腾讯混元系列模型规格。参数来源：config-center compute/providers/tencent.json。hunyuan-2.0-thinking 和 hunyuan-t1 为推理模型。",
   "specs": [
+    {
+      "id": "hy3",
+      "displayName": "混元 Hy3",
+      "family": "hy3",
+      "match": {
+        "exact": [
+          "hy3"
+        ]
+      },
+      "spec": {
+        "contextWindow": 262144,
+        "maxOutputTokens": 128000,
+        "capabilities": [
+          "chat",
+          "reasoning",
+          "code",
+          "tool_use",
+          "agent"
+        ],
+        "serviceType": [
+          "chat",
+          "reasoning"
+        ],
+        "defaultTemperature": 0.9,
+        "supportsReasoning": true,
+        "description": "腾讯混元 Hy3 正式版，295B MoE（21B active），支持可配置推理强度",
+        "releasedAt": "2026-07-06"
+      }
+    },
     {
       "id": "hy3-preview",
       "displayName": "混元 Hy3 Preview",
diff --git a/compute/model-specs/xai.json b/compute/model-specs/xai.json
index c8255e7..5a3dc44 100644
--- a/compute/model-specs/xai.json
+++ b/compute/model-specs/xai.json
@@ -1,6 +1,39 @@
 {
   "description": "xAI Grok 系列模型规格。参数来源：config-center compute/providers/xai.json。Grok 网关命名变体较多（含日期戳/-fast/-reasoning 后缀），主要依赖 family=grok-4 兜底匹配。",
   "specs": [
+    {
+      "id": "grok-4.5",
+      "displayName": "Grok 4.5",
+      "family": "grok-4.5",
+      "match": {
+        "exact": [
+          "grok-4.5"
+        ]
+      },
+      "spec": {
+        "contextWindow": 500000,
+        "maxOutputTokens": 450000,
+        "capabilities": [
+          "chat",
+          "reasoning",
+          "code",
+          "vision",
+          "long_context",
+          "tool_use"
+        ],
+        "serviceType": [
+          "chat",
+          "reasoning"
+        ],
+        "supportsReasoning": true,
+        "description": "xAI Grok 4.5，面向编程、知识工作和 STEM 的旗舰推理模型",
+        "extra": {
+          "thinkingOnly": true,
+          "thinkingDefault": true
+        },
+        "releasedAt": "2026-07-08"
+      }
+    },
     {
       "id": "grok-4-3",
       "displayName": "Grok 4.3",
diff --git a/compute/model-specs/xiaomi.json b/compute/model-specs/xiaomi.json
index 6ca248d..49f6c9f 100644
--- a/compute/model-specs/xiaomi.json
+++ b/compute/model-specs/xiaomi.json
@@ -192,6 +192,7 @@
       },
       "spec": {
         "contextWindow": 8192,
+        "maxOutputTokens": 8192,
         "capabilities": [
           "tts",
           "multilingual",
@@ -200,7 +201,43 @@
         "serviceType": [
           "tts"
         ],
-        "description": "小米 MiMo 语音合成，支持多语言与风格控制"
+        "description": "小米 MiMo 语音合成，支持多语言与风格控制",
+        "extra": {
+          "modelOrigin": {
+            "vendor": "xiaomi",
+            "canonicalModelId": "mimo-v2.5-tts",
+            "family": "mimo-tts"
+          },
+          "speech": {
+            "protocol": "xiaomi-mimo-chat-tts",
+            "protocolVersion": "1",
+            "modes": [
+              "batch",
+              "streaming"
+            ],
+            "outputFormats": [
+              "wav",
+              "mp3",
+              "pcm16le"
+            ],
+            "sampleRates": [
+              24000
+            ],
+            "defaultVoice": "mimo_default",
+            "voices": [
+              { "id": "mimo_default", "label": "MiMo 默认" },
+              { "id": "冰糖", "label": "冰糖" },
+              { "id": "茉莉", "label": "茉莉" },
+              { "id": "苏打", "label": "苏打" },
+              { "id": "白桦", "label": "白桦" },
+              { "id": "Mia", "label": "Mia" },
+              { "id": "Chloe", "label": "Chloe" },
+              { "id": "Milo", "label": "Milo" },
+              { "id": "Dean", "label": "Dean" }
+            ]
+          }
+        },
+        "releasedAt": "2026-04-23"
       },
       "routing": {
         "tier": "lightweight",
@@ -225,6 +262,8 @@
         ]
       },
       "spec": {
+        "contextWindow": 8192,
+        "maxOutputTokens": 2048,
         "capabilities": [
           "asr",
           "multilingual"
@@ -232,7 +271,27 @@
         "serviceType": [
           "asr"
         ],
-        "description": "小米 MiMo 语音识别模型"
+        "description": "小米 MiMo 语音识别模型，支持中英文语言提示与整段音频识别",
+        "extra": {
+          "modelOrigin": {
+            "vendor": "xiaomi",
+            "canonicalModelId": "mimo-v2.5-asr",
+            "family": "mimo-asr"
+          },
+          "speech": {
+            "protocol": "xiaomi-mimo-chat-asr",
+            "protocolVersion": "1",
+            "modes": [
+              "batch"
+            ],
+            "inputFormats": [
+              "wav",
+              "mp3"
+            ],
+            "languageHints": true
+          }
+        },
+        "releasedAt": "2026-06-02"
       },
       "routing": {
         "tier": "lightweight",
diff --git a/compute/providers/xiaomi.json b/compute/providers/xiaomi.json
index bae4bac..1f0d594 100644
--- a/compute/providers/xiaomi.json
+++ b/compute/providers/xiaomi.json
@@ -14,7 +14,8 @@
   "services": [
     "chat",
     "reasoning",
-    "tts"
+    "tts",
+    "asr"
   ],
   "models": [
     {
@@ -49,24 +50,37 @@
       ],
       "extra": {
         "voices": [
-          {"id": "default_zh", "label": "默认中文女声"},
-          {"id": "default_en", "label": "默认英文女声"},
           {"id": "mimo_default", "label": "MiMo 默认"},
-          {"id": "Bingtang", "label": "冰糖"},
-          {"id": "Moli", "label": "茉莉"},
-          {"id": "Suda", "label": "苏打"},
-          {"id": "Baihua", "label": "白桦"},
+          {"id": "冰糖", "label": "冰糖"},
+          {"id": "茉莉", "label": "茉莉"},
+          {"id": "苏打", "label": "苏打"},
+          {"id": "白桦", "label": "白桦"},
           {"id": "Mia", "label": "Mia"},
           {"id": "Chloe", "label": "Chloe"},
           {"id": "Milo", "label": "Milo"},
           {"id": "Dean", "label": "Dean"}
         ],
-        "audioFormats": ["mp3", "wav"],
+        "audioFormats": ["wav", "mp3", "pcm16le"],
         "sampleRate": 24000,
         "endpoint": "/chat/completions",
         "authHeader": "api-key"
       }
     },
+    {
+      "modelName": "mimo-v2.5-asr",
+      "displayName": "MiMo-V2.5-ASR",
+      "serviceType": [
+        "asr"
+      ],
+      "description": "小米 MiMo 语音识别模型，支持中英文提示与整段 WAV/MP3 识别",
+      "contextWindow": 8192,
+      "maxOutputTokens": 2048,
+      "capabilities": [
+        "asr",
+        "multilingual"
+      ],
+      "extra": {}
+    },
     {
       "modelName": "mimo-v2.5-tts-voicedesign",
       "displayName": "MiMo-V2.5-TTS-VoiceDesign",
diff --git a/manifest.json b/manifest.json
index 1e1eb10..1bdd44f 100644
--- a/manifest.json
+++ b/manifest.json
@@ -1,6 +1,6 @@
 {
   "version": "1.0.0",
-  "presetDataVersion": 106,
+  "presetDataVersion": 107,
   "updatedAt": "2026-08-31",
   "description": "DesireCore 官方配置中心"
 }
diff --git a/schemas/model-spec.schema.json b/schemas/model-spec.schema.json
index 4d50484..ba6e135 100644
--- a/schemas/model-spec.schema.json
+++ b/schemas/model-spec.schema.json
@@ -18,6 +18,107 @@
   },
   "additionalProperties": false,
   "definitions": {
+    "modelOrigin": {
+      "type": "object",
+      "required": ["vendor", "canonicalModelId"],
+      "properties": {
+        "vendor": {
+          "type": "string",
+          "minLength": 1,
+          "pattern": "^[a-z0-9]+(?:-[a-z0-9]+)*$",
+          "description": "规范化原厂商 ID，如 xiaomi、openai、alibaba、minimax；不得填写当前算力 Provider ID。"
+        },
+        "canonicalModelId": {
+          "type": "string",
+          "minLength": 1,
+          "description": "原厂模型的稳定识别 ID，用于适配器匹配、诊断和日志；不替代实际路由使用的 modelName/apiModelId。"
+        },
+        "family": {
+          "type": "string",
+          "minLength": 1,
+          "description": "跨小版本稳定的模型家族，如 mimo-asr；同一协议下的后续模型版本可复用该家族语义。"
+        }
+      },
+      "additionalProperties": false
+    },
+    "speechVoice": {
+      "type": "object",
+      "required": ["id", "label"],
+      "properties": {
+        "id": {
+          "type": "string",
+          "minLength": 1,
+          "description": "传给上游语音接口的稳定音色 ID。"
+        },
+        "label": {
+          "type": "string",
+          "minLength": 1,
+          "description": "面向用户展示的音色名称。"
+        }
+      },
+      "additionalProperties": false
+    },
+    "speechModelProfile": {
+      "type": "object",
+      "required": ["protocol", "protocolVersion", "modes"],
+      "properties": {
+        "protocol": {
+          "type": "string",
+          "minLength": 1,
+          "pattern": "^[a-z0-9]+(?:-[a-z0-9]+)*$",
+          "description": "稳定语音协议族，如 xiaomi-mimo-chat-asr；模型常规迭代不得改变此值。"
+        },
+        "protocolVersion": {
+          "type": "string",
+          "pattern": "^[1-9][0-9]*$",
+          "description": "协议不兼容版本号；只有请求、认证、响应或流式状态机发生不兼容变化时才递增。"
+        },
+        "modes": {
+          "type": "array",
+          "minItems": 1,
+          "uniqueItems": true,
+          "items": { "type": "string", "enum": ["batch", "streaming"] },
+          "description": "模型协议支持的调用模式：batch（整段）和/或 streaming（流式）。"
+        },
+        "inputFormats": {
+          "type": "array",
+          "minItems": 1,
+          "uniqueItems": true,
+          "items": { "type": "string", "minLength": 1 },
+          "description": "ASR 上游接受的音频格式，如 wav、mp3、webm；格式名使用适配器可识别的规范值。"
+        },
+        "outputFormats": {
+          "type": "array",
+          "minItems": 1,
+          "uniqueItems": true,
+          "items": { "type": "string", "minLength": 1 },
+          "description": "TTS 上游可返回的音频格式，如 pcm16le、wav、mp3；运行时需与适配器和播放器能力求交集。"
+        },
+        "sampleRates": {
+          "type": "array",
+          "minItems": 1,
+          "uniqueItems": true,
+          "items": { "type": "integer", "minimum": 1 },
+          "description": "支持的采样率（Hz）；运行时不得假设固定采样率，需完成能力协商。"
+        },
+        "defaultVoice": {
+          "type": "string",
+          "minLength": 1,
+          "description": "TTS 推荐默认音色 ID；声明 voices 时必须能在 voices 中找到同 ID 条目。"
+        },
+        "voices": {
+          "type": "array",
+          "minItems": 1,
+          "items": { "$ref": "#/definitions/speechVoice" },
+          "description": "TTS 可选音色列表，统一为稳定 id 与用户可读 label。"
+        },
+        "languageHints": {
+          "type": "boolean",
+          "description": "ASR 协议是否接受语言提示参数；不是模型支持语言范围清单。"
+        }
+      },
+      "additionalProperties": false
+    },
     "modelSpec": {
       "type": "object",
       "required": ["id", "spec"],
@@ -158,9 +259,11 @@
             },
             "extra": {
               "type": "object",
-              "description": "模型内在附加配置（如 apiModelId 全名、推理预算等）。不含价格，也不得声明 Provider/接入面的 reasoning effort 能力。",
+              "description": "模型内在开放扩展：modelOrigin 描述真实原厂身份，speech 描述可跨小版本复用的语音协议；不含价格，也不得声明 Provider/接入面的 reasoning effort 能力。",
               "not": { "required": ["reasoning"] },
               "properties": {
+                "modelOrigin": { "$ref": "#/definitions/modelOrigin" },
+                "speech": { "$ref": "#/definitions/speechModelProfile" },
                 "thinkingOnly": {
                   "type": "boolean",
                   "description": "模型强制思考、无法关闭（如 GLM-5.3 系列 thinking.type 仅接受 enabled）。true 时客户端把 off 从可选档位移除，并把内部 suppressThinking 降级为 auto，避免发出必然被上游拒绝的关闭参数。只描述模型固有属性，不声明接入面支持哪些深度档位——后者只能写在 provider model 的 extra.reasoning 中。"