AgoraIO · plutoless · Aug 3, 2026 · Jul 31, 2026
diff --git a/docs/concepts/agent.md b/docs/concepts/agent.md
@@ -65,7 +65,7 @@ Each method returns a new `Agent` instance with the updated configuration. Vendo
 | `withLlm` | `withLlm(vendor: LlmVendor): Agent<TTSSampleRate, TArea>` | Set the LLM vendor (global or CN) |
 | `withTts` | `withTts<SR>(vendor: TtsVendor<SR>): Agent<SR, TArea>` | Set the TTS vendor (tracks sample rate type) |
 | `withStt` | `withStt(vendor: SttVendor): Agent<TTSSampleRate, TArea>` | Set the STT vendor (global or CN) |
-| `withMllm` | `withMllm(vendor: BaseMLLM): Agent<TTSSampleRate, TArea>` | Set the MLLM vendor (for multimodal flow). Not compatible with `withAvatar()`. |
+| `withMllm` | `withMllm(vendor: GlobalMllmVendor \| CNMllmVendor): Agent<TTSSampleRate, TArea>` | Set the global or Chinese mainland MLLM vendor. Not compatible with `withAvatar()`. |
 | `withAvatar` | `withAvatar<SR>(this: Agent<SR, TArea>, vendor: AvatarVendor<SR>): Agent<SR, TArea>` | Set the avatar vendor (enforces TTS sample rate match). Requires the cascading pipeline; not supported with `withMllm()`. |
 | `withTurnDetection` | `withTurnDetection(config: TurnDetectionConfig): Agent<TTSSampleRate, TArea>` | Configure `turn_detection.language` and cascading-flow SOS/EOS detection |
 | `withInterruption` | `withInterruption(config: InterruptionConfig): Agent<TTSSampleRate, TArea>` | Configure unified interruption behavior |

diff --git a/docs/concepts/vendors.md b/docs/concepts/vendors.md
@@ -52,6 +52,7 @@ const llm = new OpenAI({
 | `DeepgramTTS`   | Deepgram         | Configurable                            |
 | `GradiumTTS`    | Gradium          | Configurable                            |
 | `MistralTTS`    | Mistral          | Not configurable                        |
+| `TypecastTTS`   | Typecast         | Not configurable                        |
 | `GenericTTS`    | Generic OpenAI-compatible HTTP TTS | Configurable                  |
 | `SarvamTTS`     | Sarvam AI        | Configurable                            |
 | `XAiTTS`        | xAI              | Configurable                            |
@@ -87,7 +88,7 @@ The `sampleRate` is critical when using avatars. See [Avatar Integration](../gui
 | `GoogleSTT`       | Google Speech     | `projectId`, `location`, `adcCredentialsString`, `language` |
 | `AmazonSTT`       | Amazon Transcribe | `accessKey`, `secretKey`, `region`, `language`   |
 | `AssemblyAISTT`   | AssemblyAI        | `apiKey`, `language`, `uri?`                     |
-| `AresSTT`         | Agora ARES        | —                                                |
+| `AresSTT`         | Agora ARES        | `keywords?`, `additionalParams?`                 |
 | `SarvamSTT`       | Sarvam AI         | `apiKey`, `language`                             |
 | `XAiSTT`          | xAI               | `apiKey`, `language?`, `baseUrl?`, `sampleRate?` |
 
@@ -110,9 +111,11 @@ MLLM (Multimodal LLM) vendors handle audio end-to-end — no separate STT or TTS
 | Class            | Provider                        | Key constructor params                                                                                                                                                                    |
 | ---------------- | ------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `OpenAIRealtime` | OpenAI Realtime API             | `apiKey`, `model?`, `url?`, `greetingMessage?`, `failureMessage?`, `inputModalities?`, `outputModalities?`, `messages?`, `turnDetection?`                                                 |
+| `AzureOpenAIRealtime` | Azure OpenAI Realtime (global) | `apiKey`, `url`, `turnDetection`, `model?`, `voice?`, `instructions?`, `params?`, `messages?`, `outputModalities?`, `maxHistory?`, `greetingMessage?` |
 | `GeminiLive`     | Google Gemini Live API          | `apiKey`, `model`, `url?`, `voice?`, `greetingMessage?`, `failureMessage?`, `inputModalities?`, `outputModalities?`, `messages?`, `turnDetection?`                                        |
 | `VertexAI`       | Vertex AI Gemini Live           | `model`, `url?`, `projectId`, `location`, `adcCredentialsString`, `voice?`, `greetingMessage?`, `failureMessage?`, `inputModalities?`, `outputModalities?`, `messages?`, `turnDetection?` |
 | `XaiGrok`        | xAI Grok (`mllm.vendor`: `xai`) | `apiKey`, `url?`, `voice?`, `language?`, `sampleRate?`, `greetingMessage?`, `failureMessage?`, `inputModalities?`, `outputModalities?`, `messages?`, `turnDetection?`                     |
+| `QwenOmni`       | Alibaba Cloud Qwen Omni (CN)    | `apiKey`, `model`, `turnDetection`, `url?`, `voice?`, `greetingMessage?`, `failureMessage?` |
 
 `XAiSTT` and `XAiTTS` are the cascading-pipeline xAI wrappers. `XaiGrok` is the realtime MLLM wrapper.
 
@@ -132,14 +135,15 @@ See [MLLM Flow Guide](../guides/mllm-flow.md) for full examples.
 
 ## CN vendors (Chinese mainland)
 
-CN vendor classes use the same builder methods (`.withLlm()`, `.withStt()`, `.withTts()`, `.withAvatar()`). `client.area` does not restrict which explicit vendor classes you can pass, and it also affects the default ASR vendor when `.withStt()` is omitted:
+CN vendor classes use the same builder methods (`.withLlm()`, `.withStt()`, `.withTts()`, `.withMllm()`, `.withAvatar()`). `client.area` does not restrict which explicit vendor classes you can pass, and it also affects the default ASR vendor when `.withStt()` is omitted:
 
 - `Area.CN` defaults to `FengmingSTT`
 - all other areas default to `AresSTT`
 
 | Category | Examples |
 |---|---|
 | LLM | `AliyunLLM`, `BytedanceLLM`, `DeepSeekLLM`, `TencentLLM`, `CustomLLM` |
+| MLLM | `QwenOmni` |
 | STT | `FengmingSTT`, `TencentSTT`, `MicrosoftCNSTT`, `XfyunSTT`, `XfyunBigModelSTT`, `XfyunDialectSTT` |
 | TTS | `GenericTTS` (shared with Global), `MiniMaxCNTTS`, `MicrosoftCNTTS`, `TencentTTS`, `CosyVoiceTTS`, `BytedanceDuplexTTS`, `StepFunTTS` |
 | Avatar | `SensetimeAvatar`, `SpatiusAvatar` |
@@ -150,7 +154,7 @@ See [Regional Routing](../guides/regional-routing.md) and [Vendor Reference](../
 
 Avatars provide a visual representation for the agent. Several avatar vendors require a specific TTS sample rate — this is enforced at both compile time and runtime.
 
-> Avatars currently require the cascading ASR + LLM + TTS pipeline. They are not supported with MLLM (`OpenAIRealtime`, `GeminiLive`, `VertexAI`, `XaiGrok`); combining the two throws at `Agent.toProperties()` and `AgentSession.start()`.
+> Avatars currently require the cascading ASR + LLM + TTS pipeline. They are not supported with MLLM (`OpenAIRealtime`, `AzureOpenAIRealtime`, `GeminiLive`, `VertexAI`, `XaiGrok`, `QwenOmni`); combining the two throws at `Agent.toProperties()` and `AgentSession.start()`.
 
 | Class              | Provider                                    | Required TTS sample rate |
 | ------------------ | ------------------------------------------- | ------------------------ |
@@ -162,7 +166,7 @@ Avatars provide a visual representation for the agent. Several avatar vendors re
 | `SensetimeAvatar`  | SenseTime (CN)                              | Provider-defined         |
 | `SpatiusAvatar`    | Spatius (CN)                                | Provider-defined         |
 
-CN LLM, STT, TTS, and avatar vendors (`AliyunLLM`, `FengmingSTT`, `MiniMaxCNTTS`, `SensetimeAvatar`, `SpatiusAvatar`, …) are listed in [Vendor Reference](../reference/vendors.md) and [Regional Routing](../guides/regional-routing.md).
+CN MLLM, LLM, STT, TTS, and avatar vendors (`QwenOmni`, `AliyunLLM`, `FengmingSTT`, `MiniMaxCNTTS`, `SensetimeAvatar`, `SpatiusAvatar`, …) are listed in [Vendor Reference](../reference/vendors.md) and [Regional Routing](../guides/regional-routing.md).
 
 See [Avatar Integration](../guides/avatars.md) for full examples and the sample-rate constraint details.
 

diff --git a/docs/guides/agent-builder-features.md b/docs/guides/agent-builder-features.md
@@ -349,5 +349,5 @@ const agentId = await session.start();
 
 - [Agent Reference](../reference/agent.md) — full API signatures
 - [Cascading Flow](./cascading-flow.md) — ASR → LLM → TTS setup
-- [MLLM Flow](./mllm-flow.md) — end-to-end audio with OpenAI Realtime, Gemini Live, Vertex AI, or xAI Grok
+- [MLLM Flow](./mllm-flow.md) — end-to-end audio with global providers or Chinese mainland Qwen Omni
 - [Regional Routing](./regional-routing.md) — client area and geofence
diff --git a/docs/guides/avatars.md b/docs/guides/avatars.md
@@ -20,7 +20,7 @@ Avatars attach a visual representation to the agent's audio output.
 
 ## Avatars require the cascading pipeline
 
-Avatars currently only work with the cascading **ASR + LLM + TTS** pipeline. They are **not** supported with MLLM (`OpenAIRealtime`, `GeminiLive`, `VertexAI`, `XaiGrok`). AgentKit rejects the combination at `Agent.toProperties()` and at `AgentSession.start()` so you see a clear error before the request reaches the backend:
+Avatars currently only work with the cascading **ASR + LLM + TTS** pipeline. They are **not** supported with MLLM (`OpenAIRealtime`, `AzureOpenAIRealtime`, `GeminiLive`, `VertexAI`, `XaiGrok`, `QwenOmni`). AgentKit rejects the combination at `Agent.toProperties()` and at `AgentSession.start()` so you see a clear error before the request reaches the backend:
 
 ```
 Avatars are only supported with the cascading ASR + LLM + TTS pipeline.

diff --git a/docs/guides/cascading-flow.md b/docs/guides/cascading-flow.md
@@ -197,5 +197,5 @@ await session.start();
 
 ## Next steps
 
-- [MLLM Flow](./mllm-flow.md) — skip the ASR → LLM → TTS pipeline with OpenAI Realtime, Gemini Live, Vertex AI, or xAI Grok
+- [MLLM Flow](./mllm-flow.md) — skip the ASR → LLM → TTS pipeline with a global MLLM provider or Chinese mainland Qwen Omni
 - [Avatar Integration](./avatars.md) — add a visual avatar to the agent
diff --git a/docs/guides/mllm-flow.md b/docs/guides/mllm-flow.md
@@ -1,23 +1,25 @@
 ---
 sidebar_position: 2
 title: MLLM Flow (Multimodal)
-description: Use OpenAI Realtime, Gemini Live, Vertex AI, or xAI Grok for end-to-end audio processing.
+description: Use global or Chinese mainland MLLM providers for end-to-end audio processing.
 ---
 
 # MLLM Flow (Multimodal)
 
-In MLLM mode, a single multimodal model handles audio input and output end-to-end — no separate STT or TTS step. This reduces latency and is required for OpenAI Realtime, Gemini Live, Vertex AI, and xAI Grok.
+In MLLM mode, a single multimodal model handles audio input and output end-to-end — no separate STT or TTS step. AgentKit supports global OpenAI Realtime, Azure OpenAI Realtime, Gemini Live, Vertex AI, and xAI Grok providers, plus Qwen Omni for Chinese mainland deployments.
 
 ## When to use MLLM
 
 - You want the lowest-latency conversational experience
-- You are using OpenAI Realtime API, Google Gemini Live, Vertex AI Gemini Live, or xAI Grok Realtime
+- You are using OpenAI Realtime API, Azure OpenAI Realtime, Google Gemini Live, Vertex AI Gemini Live, xAI Grok Realtime, or Qwen Omni
 - You don't need fine-grained control over STT/TTS vendor selection
 
 ## Requirements
 
 Call `agent.withMllm(vendor)` — that's it. MLLM mode is enabled automatically through `mllm.enable`. The `withLlm()`, `withTts()`, and `withStt()` methods are not needed — the MLLM vendor handles everything.
 
+`AzureOpenAIRealtime` is a global MLLM (`GlobalMllmVendor`). `QwenOmni` is a Chinese mainland MLLM (`CNMllmVendor`). As with the other explicit vendor helpers, `client.area` controls Agora REST routing but does not prevent you from selecting a provider explicitly.
+
 ## Limitations
 
 Avatars are not supported with MLLM at this time. The avatar publisher requires the cascading ASR + LLM + TTS pipeline, so combining `withMllm()` with `withAvatar()` throws at `Agent.toProperties()` and `AgentSession.start()`:
@@ -63,6 +65,31 @@ console.log('Realtime agent running:', agentId);
 await session.stop();
 ```
 
+## Example: Azure OpenAI Realtime (global)
+
+```typescript
+import { AgoraClient, Area, Agent, AzureOpenAIRealtime } from 'agora-agents';
+
+const client = new AgoraClient({
+  area: Area.US,
+  appId: 'your-app-id',
+  appCertificate: 'your-app-certificate',
+});
+
+const agent = new Agent({ client }).withMllm(new AzureOpenAIRealtime({
+  apiKey: 'your-azure-openai-key',
+  url: 'wss://example.openai.azure.com/openai/realtime',
+  params: {
+    instructions: 'You are a conversational AI agent developed by Agora.',
+    model: 'gpt-realtime-2',
+    voice: 'alloy',
+  },
+  outputModalities: ['audio'],
+  maxHistory: 32,
+  turnDetection: { mode: 'server_vad' },
+}));
+```
+
 ## Example: Gemini Live
 
 ```typescript
@@ -126,6 +153,26 @@ const agentId = await session.start();
 console.log('Grok agent running:', agentId);
 ```
 
+## Example: Qwen Omni (Chinese mainland)
+
+```typescript
+import { AgoraClient, Area, Agent, QwenOmni } from 'agora-agents';
+
+const client = new AgoraClient({
+  area: Area.CN,
+  appId: 'your-app-id',
+  appCertificate: 'your-app-certificate',
+});
+
+const agent = new Agent({ client }).withMllm(new QwenOmni({
+  apiKey: 'your-dashscope-key',
+  model: 'qwen-omni-turbo-realtime',
+  voice: 'Cherry',
+  greetingMessage: 'Hello, Qwen Omni is ready.',
+  turnDetection: { mode: 'server_vad' },
+}));
+```
+
 ## Turn detection in MLLM mode
 
 Configure MLLM turn detection on the MLLM vendor with `turnDetection`. When set, `mllm.turn_detection` overrides the top-level `turn_detection` object.

diff --git a/docs/guides/regional-routing.md b/docs/guides/regional-routing.md
@@ -35,17 +35,18 @@ const client = new AgoraClient({
 
 ## Vendor classes
 
-Import any supported vendor class and pass it to `.withStt()`, `.withLlm()`, and `.withTts()`. `client.area` does not restrict explicit provider choice, and it also determines the default ASR vendor when `.withStt()` is omitted:
+Import any supported vendor class and pass it to `.withStt()`, `.withLlm()`, `.withTts()`, or `.withMllm()`. `client.area` does not restrict explicit provider choice, and it also determines the default ASR vendor when `.withStt()` is omitted:
 
 - `Area.CN` defaults to `FengmingSTT`
 - all other areas default to `AresSTT`
 
 | Typical global providers | Typical CN providers |
 |---|---|
-| `DeepgramSTT`, `OpenAI`, `MiniMaxTTS`, `ElevenLabsTTS`, `GenericTTS`, … | `FengmingSTT`, `AliyunLLM`, `MiniMaxCNTTS`, `TencentTTS`, `GenericTTS`, … |
+| `DeepgramSTT`, `OpenAI`, `MiniMaxTTS`, `AzureOpenAIRealtime`, `GenericTTS`, … | `FengmingSTT`, `AliyunLLM`, `MiniMaxCNTTS`, `QwenOmni`, `GenericTTS`, … |
 
 See the tables below for the full catalog. You may combine any explicit vendor with any `client.area`.
 `GenericTTS` is included in both the exported `GlobalTtsVendor` and `CNTtsVendor` types.
+MLLM helpers retain their region in the exported `GlobalMllmVendor` and `CNMllmVendor` types: Azure is global and Qwen Omni is CN.
 
 Global client example:
 

diff --git a/docs/index.md b/docs/index.md
@@ -12,12 +12,12 @@ The `agora-agents` package lets you build real-time voice AI agents on the [Agor
 
 **Cascading flow** uses ASR -> LLM -> TTS and supports the broadest set of vendor combinations.
 
-**MLLM flow** uses a multimodal model such as OpenAI Realtime, Gemini Live, Vertex AI, or xAI Grok for end-to-end audio.
+**MLLM flow** uses a multimodal model such as OpenAI Realtime, Azure OpenAI Realtime, Gemini Live, Vertex AI, xAI Grok, or Qwen Omni for end-to-end audio.
 
 ## Start here
 
 - Start with [Quick Start](./getting-started/quick-start.md). It shows the baseline app-credentials setup and starts a cascading ASR -> LLM -> TTS agent.
-- Use [MLLM Flow Guide](./guides/mllm-flow.md) when your agent uses one realtime multimodal model, such as OpenAI Realtime, Gemini Live, Vertex AI, or xAI Grok.
+- Use [MLLM Flow Guide](./guides/mllm-flow.md) when your agent uses one realtime multimodal model, including global providers and Chinese mainland Qwen Omni.
 - Use [Cascading Flow Guide](./guides/cascading-flow.md) for more examples of the default ASR -> LLM -> TTS flow, including provider-specific configuration.
 
 ## How the SDK is organized
@@ -46,7 +46,7 @@ npm install agora-agents
 | [AgentSession](./concepts/session.md) | State machine, lifecycle methods, events |
 | [Vendors](./concepts/vendors.md) | LLM, TTS, STT, MLLM, and Avatar provider catalog |
 | [Cascading Flow Guide](./guides/cascading-flow.md) | Step-by-step ASR -> LLM -> TTS |
-| [MLLM Flow Guide](./guides/mllm-flow.md) | OpenAI Realtime, Gemini Live, Vertex AI, and xAI Grok |
+| [MLLM Flow Guide](./guides/mllm-flow.md) | Global MLLM providers and Chinese mainland Qwen Omni |
 | [Avatar Integration](./guides/avatars.md) | LiveAvatar, Generic Avatar, Anam, Akool, SenseTime (CN), and HeyGen integration |
 | [Agent Builder Features](./guides/agent-builder-features.md) | Turn detection, SAL, filler words, and advanced agent options |
 | [Regional Routing](./guides/regional-routing.md) | Area enum, domain pool, failover |

diff --git a/docs/reference/agent.md b/docs/reference/agent.md
@@ -78,9 +78,9 @@ Set the STT vendor. Accepts global classes (`DeepgramSTT`, `SpeechmaticsSTT`,
 
 If you omit `withStt()`, AgentKit still sends an ASR config automatically based on `client.area`: `fengming` for `Area.CN`, otherwise `ares`.
 
-### `withMllm(vendor: BaseMLLM): Agent<TTSSampleRate, TArea>`
+### `withMllm(vendor: GlobalMllmVendor | CNMllmVendor): Agent<TTSSampleRate, TArea>`
 
-Set the MLLM vendor for multimodal mode. Pass `OpenAIRealtime`, `GeminiLive`, `VertexAI`, or `XaiGrok`. Calling `withMllm()` automatically sets `mllm.enable = true`. MLLM mode does not require `withTts()` / `withLlm()` / `withStt()`.
+Set the MLLM vendor for multimodal mode. Pass a global vendor (`OpenAIRealtime`, `AzureOpenAIRealtime`, `GeminiLive`, `VertexAI`, `XaiGrok`) or the Chinese mainland vendor `QwenOmni`. Calling `withMllm()` automatically sets `mllm.enable = true`. MLLM mode does not require `withTts()` / `withLlm()` / `withStt()`.
 
 > Avatars are only supported with the cascading ASR + LLM + TTS pipeline. Combining `withMllm()` with `withAvatar()` throws at `toProperties()` and `session.start()`.