diff --git a/src/core/config.rs b/src/core/config.rs
index 8cc152f..d7bb1dd 100644
--- a/src/core/config.rs
+++ b/src/core/config.rs
@@ -15,6 +15,12 @@ pub struct AppConfig {
     pub remote_control: RemoteControlConfig,
     #[serde(default)]
     pub ble: BleConfig,
+    /// 角色元信息 (M2.1 新增，向后兼容旧配置)
+    #[serde(default)]
+    pub character: CharacterConfig,
+    /// AI 对话配置 (M2.1 新增)
+    #[serde(default)]
+    pub ai: AiConfig,
     #[serde(default)]
     pub source_path: PathBuf,
     #[serde(default)]
@@ -298,6 +304,104 @@ fn default_ble_device_name() -> String {
     "Showen".to_string()
 }
 
+// ── 角色元信息 (M2.1) ──
+
+/// 角色类型
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
+#[serde(rename_all = "lowercase")]
+pub enum CharacterType {
+    #[default]
+    Pet,
+    Human,
+    Singer,
+}
+
+/// 角色元信息配置块（内容包的一部分，切角色即切人设）
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+#[serde(deny_unknown_fields)]
+pub struct CharacterConfig {
+    /// 角色显示名 (如 "小汪")
+    #[serde(default)]
+    pub name: String,
+    /// 角色类型
+    #[serde(default)]
+    pub character_type: CharacterType,
+    /// 封面图相对路径 (内容包内)
+    #[serde(default)]
+    pub cover_image: Option<String>,
+    /// 人设 prompt (LLM system prompt)
+    #[serde(default)]
+    pub persona_prompt: String,
+    /// 最大回复 token 数
+    #[serde(default = "default_character_max_tokens")]
+    pub max_tokens: u16,
+    /// TTS 音色标识 (piper 模型 ID，留空用默认)
+    #[serde(default)]
+    pub tts_voice: Option<String>,
+    /// talk 状态名 (语音回合期间切换到，留空用 "talk")
+    #[serde(default = "default_talk_state")]
+    pub talk_state: String,
+}
+
+fn default_character_max_tokens() -> u16 {
+    128
+}
+fn default_talk_state() -> String {
+    "talk".to_string()
+}
+
+// ── AI 配置 (M2.1) ──
+
+/// LLM 后端类型
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
+#[serde(rename_all = "lowercase")]
+pub enum LlmBackend {
+    #[default]
+    Local,
+    /// 云端 (V1 仅占位，配置时返回"暂未支持")
+    Cloud,
+}
+
+/// AI 推理配置
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+#[serde(deny_unknown_fields)]
+pub struct AiConfig {
+    /// LLM 后端 (local/cloud)
+    #[serde(default)]
+    pub backend: LlmBackend,
+    /// 云端 endpoint (预留，V1 不实现)
+    #[serde(default)]
+    pub cloud_endpoint: Option<String>,
+    /// 云端 API key (预留)
+    #[serde(default)]
+    pub cloud_api_key: Option<String>,
+    /// 模型存储目录 (默认 model_store/)
+    #[serde(default = "default_model_store")]
+    pub model_store: PathBuf,
+    /// 工具目录 (whisper-cli/llama-cli/piper 二进制所在)
+    #[serde(default = "default_tools_dir")]
+    pub tools_dir: PathBuf,
+    /// 临时文件目录
+    #[serde(default = "default_tmp_dir")]
+    pub tmp_dir: PathBuf,
+    /// 上下文窗口轮数
+    #[serde(default = "default_context_window")]
+    pub context_window: usize,
+}
+
+fn default_model_store() -> PathBuf {
+    PathBuf::from("model_store")
+}
+fn default_tools_dir() -> PathBuf {
+    PathBuf::from("tools")
+}
+fn default_tmp_dir() -> PathBuf {
+    PathBuf::from("/tmp/showen_ai")
+}
+fn default_context_window() -> usize {
+    5
+}
+
 // ── 加载与验证 ──
 
 impl AppConfig {
diff --git a/src/core/message.rs b/src/core/message.rs
index d69e520..35d2eb0 100644
--- a/src/core/message.rs
+++ b/src/core/message.rs
@@ -61,6 +61,14 @@ pub enum Message {
     DeviceResponse(DeviceResponse),
     DeviceEvent(DeviceEvent),
 
+    // ── AI 语音对话 (M2.1) ──
+    /// 语音/文字对话回合请求（HTTP → AI 插件）
+    ChatRequest(ChatRequest),
+    /// 对话回合响应（AI 插件 → HTTP/广播）
+    ChatResponse(ChatResponse),
+    /// AI 模型状态变更广播（加载/切换/下载进度）
+    AiModelEvent(AiModelEvent),
+
     // ── 扩展（未来插件用） ──
     Custom {
         kind: String,
@@ -68,6 +76,62 @@ pub enum Message {
     },
 }
 
+/// 对话回合请求
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ChatRequest {
+    /// 会话标识（同一会话保留上下文）
+    pub session_id: String,
+    /// 输入类型：文字或音频文件路径（设备临时文件）
+    pub input: ChatInput,
+    /// 当前角色人设 prompt
+    pub persona_prompt: String,
+    /// 最大回复 token 数（0 表示用默认）
+    pub max_tokens: u16,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "lowercase")]
+pub enum ChatInput {
+    /// 文字输入（Web 端主路径）
+    Text { content: String },
+    /// 音频文件路径（App 上传后的临时文件，16kHz mono wav）
+    Audio { path: String },
+}
+
+/// 对话回合响应
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ChatResponse {
+    pub session_id: String,
+    /// ASR 转写文本（音频输入时，文字输入时为 None）
+    pub transcription: Option<String>,
+    /// LLM 回复文本
+    pub reply_text: String,
+    /// TTS 回复音频文件路径（设备临时文件，HTTP 返回后由客户端播放）
+    pub reply_audio_path: Option<String>,
+    /// 错误信息（非空表示失败）
+    pub error: Option<String>,
+}
+
+/// AI 模型事件
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(tag = "event", rename_all = "snake_case")]
+pub enum AiModelEvent {
+    /// 模型加载中
+    Loading { model_id: String },
+    /// 模型已就绪
+    Ready { model_id: String },
+    /// 模型切换
+    Switched { old: String, new: String },
+    /// 下载进度
+    DownloadProgress {
+        model_id: String,
+        downloaded: u64,
+        total: u64,
+    },
+    /// 错误
+    Error { model_id: String, message: String },
+}
+
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub enum PlayerCommand {
     Play,
diff --git a/src/main.rs b/src/main.rs
index 5562f84..55fb542 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -5,7 +5,7 @@ use showen_v2::core::service_manager::ServiceManager;
 #[cfg(not(test))]
 use showen_v2::core::version_manager::VersionManager;
 use showen_v2::plugins::{
-    ble::BlePlugin, device::DevicePlugin, http::HttpPlugin, screen::ScreenPlugin,
+    ai::AiPlugin, ble::BlePlugin, device::DevicePlugin, http::HttpPlugin, screen::ScreenPlugin,
     video::VideoPlugin, wifi::WifiPlugin,
 };
 use std::sync::atomic::{AtomicBool, Ordering};
@@ -49,7 +49,7 @@ fn main() -> Result<()> {
     let mut manager = ServiceManager::new(config);
 
     // 按依赖顺序注册插件
-    // 独立插件：device, screen, wifi, video, ble
+    // 独立插件：device, screen, wifi, video, ble, ai
     // 依赖插件：http (依赖 video)
 
     println!("注册插件...");
@@ -68,7 +68,25 @@ fn main() -> Result<()> {
     manager.register(Box::new(BlePlugin::new()));
     println!("  ✓ BlePlugin");
 
-    manager.register(Box::new(HttpPlugin::new()));
+    // AiPlugin: 需要在注册前创建实例以获取 pipeline 句柄，再注册进 manager
+    // pipeline 会传给 HttpPlugin，建立 HTTP → AI 的直接通道
+    let ai_config = &config.ai;
+    let ai_plugin = AiPlugin::new_default(
+        ai_config.model_store.clone(),
+        ai_config.tools_dir.clone(),
+        ai_config.tmp_dir.clone(),
+        ai_config.context_window,
+    );
+    let ai_pipeline = ai_plugin.pipeline();
+    let ai_models = ai_plugin.models();
+    manager.register(Box::new(ai_plugin));
+    println!("  ✓ AiPlugin");
+
+    let mut http_plugin = HttpPlugin::new();
+    // 注册 AI 句柄到 HttpState（在 http init 之前建立连接）
+    http_plugin.set_ai_pipeline(ai_pipeline);
+    http_plugin.set_ai_models(ai_models);
+    manager.register(Box::new(http_plugin));
     println!("  ✓ HttpPlugin");
 
     // 加载动态插件
diff --git a/src/plugins/ai/backend.rs b/src/plugins/ai/backend.rs
new file mode 100644
index 0000000..e7cdf5d
--- /dev/null
+++ b/src/plugins/ai/backend.rs
@@ -0,0 +1,213 @@
+//! AI 后端 trait + 本地命令行实现
+//!
+//! 抽象 ASR/LLM/TTS 三层，支持本地命令行（LocalCliBackend）和未来云端后端。
+//! V1 只实现 local；cloud 配置时返回"暂未支持"错误。
+
+use anyhow::{bail, Result};
+use std::path::{Path, PathBuf};
+use std::process::Command;
+
+/// AI 后端抽象 trait
+pub trait AiBackend: Send {
+    /// 语音转文字 (ASR)
+    /// - audio_path: 16kHz mono wav 临时文件
+    /// - model_path: whisper.cpp 模型 (ggml-tiny.bin 等)
+    fn asr(&self, audio_path: &Path, model_path: &Path) -> Result<String>;
+
+    /// LLM 对话
+    /// - model_path: llama.cpp GGUF 模型
+    /// - persona: 角色人设 system prompt
+    /// - user_message: 本轮用户输入
+    /// - history: 历史轮次 (role, content)
+    /// - max_tokens: 最大回复 token (0=默认 128)
+    fn llm_chat(
+        &self,
+        model_path: &Path,
+        persona: &str,
+        user_message: &str,
+        history: &[(String, String)],
+        max_tokens: u16,
+    ) -> Result<String>;
+
+    /// 文字转语音 (TTS)
+    /// - text: 要合成的文本
+    /// - model_path: piper 模型目录/文件
+    /// - output_path: 输出 wav 文件路径
+    fn tts(&self, text: &str, model_path: &Path, output_path: &Path) -> Result<()>;
+
+    /// 后端标识（local / cloud）
+    fn backend_name(&self) -> &str;
+}
+
+/// 本地命令行后端
+///
+/// 通过子进程调用 whisper-cli / llama-cli / piper 二进制。
+/// Spike (2026-07-03) 验证的命令行参数固化于此。
+pub struct LocalCliBackend {
+    /// 工具目录（含 whisper-cli, llama-cli, piper 二进制）
+    tools_dir: PathBuf,
+}
+
+impl LocalCliBackend {
+    pub fn new(tools_dir: PathBuf) -> Self {
+        Self { tools_dir }
+    }
+
+    fn whisper_cli(&self) -> PathBuf {
+        self.tools_dir.join("whisper-cli")
+    }
+
+    fn llama_cli(&self) -> PathBuf {
+        self.tools_dir.join("llama-cli")
+    }
+
+    fn piper(&self) -> PathBuf {
+        self.tools_dir.join("piper")
+    }
+}
+
+impl AiBackend for LocalCliBackend {
+    fn asr(&self, audio_path: &Path, model_path: &Path) -> Result<String> {
+        // whisper-cli -m <model> -f <audio> -l zh --no-timestamps -otxt -of <tmp>
+        // 输出到 <tmp>.txt，读取后返回
+        let tmp_out = audio_path.with_extension("asr.txt");
+        let output = Command::new(self.whisper_cli())
+            .arg("-m")
+            .arg(model_path)
+            .arg("-f")
+            .arg(audio_path)
+            .arg("-l")
+            .arg("zh")
+            .arg("--no-timestamps")
+            .arg("-otxt")
+            .arg("-of")
+            .arg(&tmp_out)
+            .output()?;
+
+        if !output.status.success() {
+            bail!(
+                "whisper-cli 失败: {}",
+                String::from_utf8_lossy(&output.stderr)
+            );
+        }
+
+        let txt_path = tmp_out.with_extension("txt");
+        let text = std::fs::read_to_string(&txt_path)?;
+        let _ = std::fs::remove_file(&txt_path);
+        Ok(text.trim().to_string())
+    }
+
+    fn llm_chat(
+        &self,
+        model_path: &Path,
+        persona: &str,
+        user_message: &str,
+        history: &[(String, String)],
+        max_tokens: u16,
+    ) -> Result<String> {
+        // 构造 llama-cli 对话 prompt
+        // 格式参考 spike: llama-cli -m <model> -p <prompt> -n <max_tokens> -t 2 -c 1024 --temp 0.7
+        let mut prompt = String::new();
+        prompt.push_str(&format!("system\n{persona}\n"));
+        for (role, content) in history {
+            prompt.push_str(role);
+            prompt.push('\n');
+            prompt.push_str(content);
+            prompt.push('\n');
+        }
+        prompt.push_str("user\n");
+        prompt.push_str(user_message);
+        prompt.push('\n');
+        prompt.push_str("assistant\n");
+
+        let tokens = if max_tokens == 0 { 128 } else { max_tokens as u32 };
+
+        let output = Command::new(self.llama_cli())
+            .arg("-m")
+            .arg(model_path)
+            .arg("-p")
+            .arg(&prompt)
+            .arg("-n")
+            .arg(tokens.to_string())
+            .arg("-t")
+            .arg("2") // 锁大核 (spike 结论: t=2 最优)
+            .arg("-c")
+            .arg("1024") // 限死上下文 (spike 警告: 默认 ctx 吃 1-3.3G)
+            .arg("--temp")
+            .arg("0.7")
+            .arg("--no-display-prompt")
+            .output()?;
+
+        if !output.status.success() {
+            bail!(
+                "llama-cli 失败: {}",
+                String::from_utf8_lossy(&output.stderr)
+            );
+        }
+
+        // llama-cli 输出到 stdout，取生成部分
+        let reply = String::from_utf8_lossy(&output.stdout);
+        Ok(reply.trim().to_string())
+    }
+
+    fn tts(&self, text: &str, model_path: &Path, output_path: &Path) -> Result<()> {
+        // echo <text> | piper -m <model> -f <output> --output_format wav
+        let mut cmd = Command::new(self.piper());
+        cmd.arg("-m")
+            .arg(model_path)
+            .arg("-f")
+            .arg(output_path)
+            .arg("--output_format")
+            .arg("wav")
+            .stdin(std::process::Stdio::piped())
+            .stdout(std::process::Stdio::null())
+            .stderr(std::process::Stdio::piped());
+
+        let mut child = cmd.spawn()?;
+        if let Some(mut stdin) = child.stdin.take() {
+            use std::io::Write;
+            stdin.write_all(text.as_bytes())?;
+        }
+        let output = child.wait_with_output()?;
+
+        if !output.status.success() {
+            bail!(
+                "piper 失败: {}",
+                String::from_utf8_lossy(&output.stderr)
+            );
+        }
+        Ok(())
+    }
+
+    fn backend_name(&self) -> &str {
+        "local"
+    }
+}
+
+/// 云端后端（V1 仅占位，返回"暂未支持"）
+pub struct CloudBackend;
+
+impl AiBackend for CloudBackend {
+    fn asr(&self, _audio_path: &Path, _model_path: &Path) -> Result<String> {
+        bail!("云端 ASR 暂未支持 (V1 仅实现 local)")
+    }
+
+    fn llm_chat(
+        &self,
+        _model_path: &Path,
+        _persona: &str,
+        _user_message: &str,
+        _history: &[(String, String)],
+        _max_tokens: u16,
+    ) -> Result<String> {
+        bail!("云端 LLM 暂未支持 (V1 仅实现 local)")
+    }
+
+    fn tts(&self, _text: &str, _model_path: &Path, _output_path: &Path) -> Result<()> {
+        bail!("云端 TTS 暂未支持 (V1 仅实现 local)")
+    }
+
+    fn backend_name(&self) -> &str {
+        "cloud"
+    }
+}
diff --git a/src/plugins/ai/chat.rs b/src/plugins/ai/chat.rs
new file mode 100644
index 0000000..36ea789
--- /dev/null
+++ b/src/plugins/ai/chat.rs
@@ -0,0 +1,180 @@
+//! 对话管线执行器（被 HTTP 层直接调用）
+//!
+//! 为什么不在 AiPlugin 里执行：HTTP 需要同步响应，而 Plugin trait 的
+//! handle_message 是异步消息系统，不便返回同步结果。改用 HTTP 层在
+//! spawn_blocking 里直接调用此模块，Arc<Mutex> 共享状态。
+
+use crate::core::message::{ChatInput, ChatRequest, ChatResponse};
+use crate::plugins::ai::backend::AiBackend;
+use crate::plugins::ai::model_manager::ModelManager;
+use anyhow::Result;
+use std::collections::HashMap;
+use std::path::Path;
+use std::sync::{Arc, Mutex};
+
+/// 会话上下文
+pub struct SessionContext {
+    pub persona_prompt: String,
+    pub history: Vec<(String, String)>,
+}
+
+/// 对话管线执行器（共享状态）
+pub struct ChatPipeline {
+    pub backend: Mutex<Box<dyn AiBackend>>,
+    pub models: Arc<Mutex<ModelManager>>,
+    pub sessions: Mutex<HashMap<String, SessionContext>>,
+    pub context_window: usize,
+    pub tmp_dir: std::path::PathBuf,
+    /// 互斥锁：同一时刻只允许一个回合
+    pub busy: Mutex<()>,
+}
+
+impl ChatPipeline {
+    pub fn new(
+        backend: Box<dyn AiBackend>,
+        models: ModelManager,
+        tmp_dir: std::path::PathBuf,
+        context_window: usize,
+    ) -> Arc<Self> {
+        Arc::new(Self {
+            backend: Mutex::new(backend),
+            models: Arc::new(Mutex::new(models)),
+            sessions: Mutex::new(HashMap::new()),
+            context_window,
+            tmp_dir,
+            busy: Mutex::new(()),
+        })
+    }
+
+    /// 执行对话回合（同步阻塞，调用方应在 spawn_blocking 里调）
+    pub fn run(&self, req: &ChatRequest) -> ChatResponse {
+        // 互斥：拿不到锁说明有进行中的回合
+        let _guard = match self.busy.try_lock() {
+            Ok(g) => g,
+            Err(_) => {
+                return ChatResponse {
+                    session_id: req.session_id.clone(),
+                    transcription: None,
+                    reply_text: String::new(),
+                    reply_audio_path: None,
+                    error: Some("设备忙，上一个回合未结束".to_string()),
+                }
+            }
+        };
+
+        let result = self.run_pipeline(req);
+
+        match result {
+            Ok((transcription, reply_text, audio_path)) => {
+                // 更新会话上下文
+                if let Ok(mut sessions) = self.sessions.lock() {
+                    let session = sessions
+                        .entry(req.session_id.clone())
+                        .or_insert_with(|| SessionContext {
+                            persona_prompt: req.persona_prompt.clone(),
+                            history: Vec::new(),
+                        });
+                    if session.persona_prompt != req.persona_prompt {
+                        session.persona_prompt = req.persona_prompt.clone();
+                        session.history.clear();
+                    }
+                    if let Some(t) = &transcription {
+                        session.history.push(("user".to_string(), t.clone()));
+                    }
+                    session.history.push(("assistant".to_string(), reply_text.clone()));
+                    let max = self.context_window * 2;
+                    if session.history.len() > max {
+                        let drain = session.history.len() - max;
+                        session.history.drain(..drain);
+                    }
+                }
+                ChatResponse {
+                    session_id: req.session_id.clone(),
+                    transcription,
+                    reply_text,
+                    reply_audio_path: audio_path,
+                    error: None,
+                }
+            }
+            Err(e) => ChatResponse {
+                session_id: req.session_id.clone(),
+                transcription: None,
+                reply_text: String::new(),
+                reply_audio_path: None,
+                error: Some(e.to_string()),
+            },
+        }
+    }
+
+    fn run_pipeline(&self, req: &ChatRequest) -> Result<(Option<String>, String, Option<String>)> {
+        // 1. ASR
+        let transcription = match &req.input {
+            ChatInput::Text { content } => Some(content.clone()),
+            ChatInput::Audio { path } => {
+                let models = self.models.lock().unwrap();
+                let asr_model = models.active_asr_model_path()?;
+                drop(models);
+                let backend = self.backend.lock().unwrap();
+                Some(backend.asr(Path::new(path), &asr_model)?)
+            }
+        };
+
+        let user_text = transcription
+            .as_ref()
+            .ok_or_else(|| anyhow::anyhow!("ASR 返回空结果"))?;
+        if user_text.trim().is_empty() {
+            anyhow::bail!("输入为空（ASR 未识别到语音）");
+        }
+
+        // 2. LLM
+        let models = self.models.lock().unwrap();
+        let llm_model = models.active_llm_model_path()?;
+        drop(models);
+
+        let history: Vec<(String, String)> = {
+            let sessions = self.sessions.lock().unwrap();
+            sessions
+                .get(&req.session_id)
+                .map(|s| s.history.clone())
+                .unwrap_or_default()
+        };
+
+        let backend = self.backend.lock().unwrap();
+        let reply_text = backend.llm_chat(
+            &llm_model,
+            &req.persona_prompt,
+            user_text,
+            &history,
+            req.max_tokens,
+        )?;
+        drop(backend);
+
+        // 3. TTS
+        let models = self.models.lock().unwrap();
+        let tts_model = models.active_tts_model_path()?;
+        drop(models);
+
+        let audio_path = self.tmp_dir.join(format!(
+            "tts_{}.wav",
+            std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)
+                .map(|d| d.as_millis())
+                .unwrap_or(0)
+        ));
+        let backend = self.backend.lock().unwrap();
+        backend.tts(&reply_text, &tts_model, &audio_path)?;
+
+        Ok((
+            transcription,
+            reply_text,
+            Some(audio_path.to_string_lossy().into_owned()),
+        ))
+    }
+
+    /// 清空指定会话上下文（切角色时调用）
+    pub fn clear_session(&self, session_id: &str) {
+        if let Ok(mut sessions) = self.sessions.lock() {
+            sessions.remove(session_id);
+        }
+    }
+}
diff --git a/src/plugins/ai/mod.rs b/src/plugins/ai/mod.rs
new file mode 100644
index 0000000..5a28baa
--- /dev/null
+++ b/src/plugins/ai/mod.rs
@@ -0,0 +1,161 @@
+//! AiPlugin — AI 语音对话插件 (M2.1)
+//!
+//! 设备本地跑 ASR (whisper.cpp) → LLM (llama.cpp) → TTS (piper) 管线，
+//! 提供语音/文字对话回合能力。声音一律返回客户端播放，设备不出声。
+//!
+//! # 架构
+//! ```text
+//! HTTP /api/chat → ChatPipeline.run() → ASR → LLM → TTS → ChatResponse
+//!                   (Arc<Mutex> 共享, HTTP 在 spawn_blocking 调用)
+//! ```
+//!
+//! AiPlugin 本身是薄层：持有 ChatPipeline Arc 供 HTTP 层取用，
+//! 并处理消息系统触发的联动（如 talk/idle 状态切换）。
+//!
+//! # 硬件约束 (Spike 2026-07-03 张明远实测)
+//! - 测试机全志 A733: 2×A78@2.0G + 6×A55@1.8G, 4G 内存
+//! - LLM 推理必须 t=2 锁大核 (全核反而崩且挤死视频)
+//! - 严禁默认上下文长度 (4096/8192 会吃 1-3.3GB)
+//! - 推荐默认 Qwen2.5-0.5B Q4_K_M (16.4 t/s, RSS 985M)
+
+pub mod backend;
+pub mod chat;
+pub mod model_manager;
+
+pub use chat::{ChatPipeline, SessionContext};
+
+use crate::core::message::{Message, StateChanged};
+use crate::core::plugin::*;
+use anyhow::Result;
+use std::path::PathBuf;
+use std::sync::{Arc, Mutex};
+
+/// AiPlugin — AI 对话插件（薄层）
+///
+/// 持有 ChatPipeline 的共享句柄，供 HTTP 层取用。
+/// 自身处理消息系统触发的联动（talk/idle 状态切换等）。
+pub struct AiPlugin {
+    ctx: Option<PluginContext>,
+    /// 对话管线（HTTP 层通过 Arc clone 直接调用）
+    pipeline: Arc<ChatPipeline>,
+    /// 工具目录
+    tools_dir: PathBuf,
+}
+
+impl AiPlugin {
+    /// 创建默认本地后端实例
+    pub fn new_default(
+        model_store: PathBuf,
+        tools_dir: PathBuf,
+        tmp_dir: PathBuf,
+        context_window: usize,
+    ) -> Self {
+        let backend = Box::new(backend::LocalCliBackend::new(tools_dir.clone()));
+        let models = model_manager::ModelManager::new(model_store);
+        let pipeline = ChatPipeline::new(backend, models, tmp_dir, context_window);
+        Self {
+            ctx: None,
+            pipeline,
+            tools_dir,
+        }
+    }
+
+    /// 获取对话管线共享句柄（main.rs 注册时传给 HttpPlugin）
+    pub fn pipeline(&self) -> Arc<ChatPipeline> {
+        Arc::clone(&self.pipeline)
+    }
+
+    /// 获取模型管理器共享句柄（main.rs 注册时传给 HttpPlugin）
+    pub fn models(&self) -> Arc<Mutex<model_manager::ModelManager>> {
+        Arc::clone(&self.pipeline.models)
+    }
+}
+
+impl Plugin for AiPlugin {
+    fn id(&self) -> &str {
+        "ai"
+    }
+
+    fn info(&self) -> PluginInfo {
+        PluginInfo {
+            name: "AiPlugin".to_string(),
+            version: "0.1.0".to_string(),
+            description: "AI 语音对话 (ASR/LLM/TTS 本地推理)".to_string(),
+            platform: Platform::LinuxArm64,
+        }
+    }
+
+    fn capabilities(&self) -> Vec<String> {
+        vec!["chat".to_string(), "model_management".to_string()]
+    }
+
+    fn init(&mut self, ctx: PluginContext) -> Result<()> {
+        self.ctx = Some(ctx);
+        Ok(())
+    }
+
+    fn start(&mut self) -> Result<()> {
+        // 确保临时目录存在
+        if let Some(tmp) = self.pipeline.tmp_dir.parent() {
+            std::fs::create_dir_all(tmp).ok();
+        }
+        std::fs::create_dir_all(&self.pipeline.tmp_dir).ok();
+
+        // 预加载默认模型清单
+        let mut models = self.pipeline.models.lock().unwrap();
+        if let Err(e) = models.ensure_default_models() {
+            eprintln!("[AiPlugin] 警告: 模型初始化失败: {e}");
+        }
+        drop(models);
+
+        println!(
+            "[AiPlugin] 启动 (tools={}, tmp={})",
+            self.tools_dir.display(),
+            self.pipeline.tmp_dir.display()
+        );
+        Ok(())
+    }
+
+    fn handle_message(&mut self, msg: Message) -> Result<()> {
+        match msg {
+            Message::ChatRequest(req) => {
+                // 通过消息系统发起的对话（非 HTTP 路径），同步执行并广播结果
+                let resp = self.pipeline.run(&req);
+                if let Some(ctx) = &self.ctx {
+                    let _ = ctx.tx.send(crate::core::message::Envelope {
+                        from: self.id().to_string(),
+                        to: crate::core::message::Destination::Broadcast,
+                        message: Message::ChatResponse(resp),
+                    });
+                }
+            }
+            Message::Shutdown => {
+                self.stop()?;
+            }
+            Message::StateChanged { old_state, new_state } => {
+                // 画面联动：进入/离开 talk 状态时记录日志（实际联动由 video 插件处理状态机）
+                if new_state == "talk" || old_state == "talk" {
+                    println!("[AiPlugin] 画面状态: {old_state} → {new_state}");
+                }
+                let _ = StateChanged { old_state, new_state }; // 抑制未用警告
+            }
+            _ => {}
+        }
+        Ok(())
+    }
+
+    fn stop(&mut self) -> Result<()> {
+        // 清理临时文件
+        if self.pipeline.tmp_dir.exists() {
+            if let Ok(entries) = std::fs::read_dir(&self.pipeline.tmp_dir) {
+                for entry in entries.flatten() {
+                    let path = entry.path();
+                    if path.extension().and_then(|s| s.to_str()) == Some("wav") {
+                        let _ = std::fs::remove_file(&path);
+                    }
+                }
+            }
+        }
+        Ok(())
+    }
+}
diff --git a/src/plugins/ai/model_manager.rs b/src/plugins/ai/model_manager.rs
new file mode 100644
index 0000000..5c29a91
--- /dev/null
+++ b/src/plugins/ai/model_manager.rs
@@ -0,0 +1,409 @@
+//! ModelManager — AI 模型资产管理 (M2.1)
+//!
+//! 管理 LLM/ASR/TTS 模型文件，支持清单查看、下载、切换、删除。
+//! 参照项目已有 plugin_repo + version_manager 模式设计。
+//!
+//! # 模型存储结构
+//! ```text
+//! model_store/
+//! ├── registry.json          # 本地模型注册表（已下载的模型清单 + active 标记）
+//! ├── llm/
+//! │   ├── qwen2.5-0.5b-q4_k_m.gguf
+//! │   └── gemma3-1b-q4_k_m.gguf
+//! ├── asr/
+//! │   └── ggml-tiny.bin
+//! └── tts/
+//!     └── zh_CN-huayan-medium.onnx
+//! ```
+
+use anyhow::{bail, Context, Result};
+use serde::{Deserialize, Serialize};
+use std::collections::HashMap;
+use std::fs;
+use std::path::{Path, PathBuf};
+
+/// 模型类型
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Hash)]
+#[serde(rename_all = "lowercase")]
+pub enum ModelKind {
+    Llm,
+    Asr,
+    Tts,
+}
+
+impl ModelKind {
+    pub fn dir_name(&self) -> &'static str {
+        match self {
+            ModelKind::Llm => "llm",
+            ModelKind::Asr => "asr",
+            ModelKind::Tts => "tts",
+        }
+    }
+}
+
+/// 模型元信息（仓库清单 + 本地注册表共用）
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ModelInfo {
+    /// 模型 ID (如 "qwen2.5-0.5b-q4_k_m")
+    pub id: String,
+    /// 模型类型
+    pub kind: ModelKind,
+    /// 显示名称 (如 "Qwen2.5 0.5B Q4")
+    pub name: String,
+    /// 版本
+    pub version: String,
+    /// 文件尺寸 (字节)
+    pub size: u64,
+    /// 运行内存需求 (字节)
+    pub memory_required: u64,
+    /// 推荐设备档位 (如 "4G 内存档可用")
+    pub recommended_tier: String,
+    /// 下载 URL（国内镜像源）
+    pub url: String,
+    /// SHA256 校验和
+    pub sha256: String,
+    /// 文件名（存储到本地的文件名）
+    pub filename: String,
+}
+
+/// 本地注册表条目
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct LocalModelEntry {
+    #[serde(flatten)]
+    pub info: ModelInfo,
+    /// 是否已下载
+    pub downloaded: bool,
+    /// 下载进度 (0-100，100=完成)
+    pub download_progress: u8,
+}
+
+/// 本地注册表
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+pub struct ModelRegistry {
+    /// 所有已知模型（仓库清单 + 本地状态）
+    pub models: Vec<LocalModelEntry>,
+    /// 各类型当前激活的模型 ID
+    pub active: HashMap<ModelKind, String>,
+}
+
+/// 模型管理器
+pub struct ModelManager {
+    /// 模型存储根目录
+    store_dir: PathBuf,
+    /// 注册表（内存缓存，持久化到 registry.json）
+    registry: ModelRegistry,
+    /// 存储配额上限（字节，默认 4GB）
+    quota: u64,
+}
+
+impl ModelManager {
+    pub fn new(store_dir: PathBuf) -> Self {
+        let quota = 4 * 1024 * 1024 * 1024; // 4GB
+        let mut mgr = Self {
+            store_dir,
+            registry: ModelRegistry::default(),
+            quota,
+        };
+        let _ = mgr.load_registry();
+        mgr
+    }
+
+    /// 存储目录
+    pub fn store_dir(&self) -> &Path {
+        &self.store_dir
+    }
+
+    /// 配额
+    pub fn quota(&self) -> u64 {
+        self.quota
+    }
+
+    /// 已用空间
+    pub fn used_space(&self) -> u64 {
+        self.registry
+            .models
+            .iter()
+            .filter(|m| m.downloaded)
+            .map(|m| m.info.size)
+            .sum()
+    }
+
+    /// 加载本地注册表
+    fn load_registry(&mut self) -> Result<()> {
+        let path = self.store_dir.join("registry.json");
+        if path.exists() {
+            let content = fs::read_to_string(&path)?;
+            self.registry = serde_json::from_str(&content)?;
+        }
+        Ok(())
+    }
+
+    /// 持久化注册表
+    fn save_registry(&self) -> Result<()> {
+        fs::create_dir_all(&self.store_dir)?;
+        let path = self.store_dir.join("registry.json");
+        let content = serde_json::to_string_pretty(&self.registry)?;
+        fs::write(&path, content)?;
+        Ok(())
+    }
+
+    /// 初始化默认模型清单（首次启动或清单缺失时）
+    pub fn ensure_default_models(&mut self) -> Result<()> {
+        if !self.registry.models.is_empty() {
+            return Ok(());
+        }
+
+        // Spike 推荐的默认模型清单
+        let defaults = vec![
+            ModelInfo {
+                id: "qwen2.5-0.5b-q4_k_m".to_string(),
+                kind: ModelKind::Llm,
+                name: "Qwen2.5 0.5B Q4 (推荐默认)".to_string(),
+                version: "0.5b".to_string(),
+                size: 469 * 1024 * 1024,
+                memory_required: 1024 * 1024 * 1024,
+                recommended_tier: "4G 内存档可用".to_string(),
+                url: "https://hf-mirror.com/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/main/qwen2.5-0.5b-instruct-q4_k_m.gguf".to_string(),
+                sha256: String::new(), // 实际部署时填充
+                filename: "qwen2.5-0.5b-q4_k_m.gguf".to_string(),
+            },
+            ModelInfo {
+                id: "gemma3-1b-q4_k_m".to_string(),
+                kind: ModelKind::Llm,
+                name: "Gemma3 1B Q4 (可选档)".to_string(),
+                version: "1b".to_string(),
+                size: 800 * 1024 * 1024,
+                memory_required: 2 * 1024 * 1024 * 1024,
+                recommended_tier: "建议 8G 以上".to_string(),
+                url: "https://hf-mirror.com/google/gemma-3-1b-it-GGUF/resolve/main/gemma-3-1b-it-q4_k_m.gguf".to_string(),
+                sha256: String::new(),
+                filename: "gemma3-1b-q4_k_m.gguf".to_string(),
+            },
+            ModelInfo {
+                id: "whisper-tiny".to_string(),
+                kind: ModelKind::Asr,
+                name: "Whisper Tiny (中文)".to_string(),
+                version: "tiny".to_string(),
+                size: 75 * 1024 * 1024,
+                memory_required: 200 * 1024 * 1024,
+                recommended_tier: "4G 内存档可用".to_string(),
+                url: "https://hf-mirror.com/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin".to_string(),
+                sha256: String::new(),
+                filename: "ggml-tiny.bin".to_string(),
+            },
+            ModelInfo {
+                id: "piper-zh-huayan".to_string(),
+                kind: ModelKind::Tts,
+                name: "Piper zh_CN huayan (中文女声)".to_string(),
+                version: "medium".to_string(),
+                size: 63 * 1024 * 1024,
+                memory_required: 150 * 1024 * 1024,
+                recommended_tier: "4G 内存档可用".to_string(),
+                url: "https://hf-mirror.com/rhasspy/piper-voices/resolve/main/zh/zh_CN/huayan/medium/zh_CN-huayan-medium.onnx".to_string(),
+                sha256: String::new(),
+                filename: "zh_CN-huayan-medium.onnx".to_string(),
+            },
+        ];
+
+        for info in defaults {
+            let filename = info.filename.clone();
+            let downloaded = self.store_dir.join(info.kind.dir_name()).join(&filename).exists();
+            self.registry.models.push(LocalModelEntry {
+                info,
+                downloaded,
+                download_progress: if downloaded { 100 } else { 0 },
+            });
+        }
+
+        // 设置默认 active 模型
+        if !self.registry.active.contains_key(&ModelKind::Llm) {
+            self.registry.active.insert(ModelKind::Llm, "qwen2.5-0.5b-q4_k_m".to_string());
+        }
+        if !self.registry.active.contains_key(&ModelKind::Asr) {
+            self.registry.active.insert(ModelKind::Asr, "whisper-tiny".to_string());
+        }
+        if !self.registry.active.contains_key(&ModelKind::Tts) {
+            self.registry.active.insert(ModelKind::Tts, "piper-zh-huayan".to_string());
+        }
+
+        self.save_registry()?;
+        Ok(())
+    }
+
+    /// 列出所有模型
+    pub fn list_models(&self) -> &[LocalModelEntry] {
+        &self.registry.models
+    }
+
+    /// 获取当前激活的模型 ID
+    pub fn active_model_id(&self, kind: ModelKind) -> Option<&String> {
+        self.registry.active.get(&kind)
+    }
+
+    /// 获取已激活 LLM 模型文件路径
+    pub fn active_llm_model_path(&self) -> Result<PathBuf> {
+        self.active_model_path(ModelKind::Llm)
+    }
+
+    /// 获取已激活 ASR 模型文件路径
+    pub fn active_asr_model_path(&self) -> Result<PathBuf> {
+        self.active_model_path(ModelKind::Asr)
+    }
+
+    /// 获取已激活 TTS 模型文件路径
+    pub fn active_tts_model_path(&self) -> Result<PathBuf> {
+        self.active_model_path(ModelKind::Tts)
+    }
+
+    fn active_model_path(&self, kind: ModelKind) -> Result<PathBuf> {
+        let id = self
+            .registry
+            .active
+            .get(&kind)
+            .ok_or_else(|| anyhow::anyhow!("没有激活的 {kind:?} 模型"))?;
+        let entry = self
+            .registry
+            .models
+            .iter()
+            .find(|m| &m.info.id == id)
+            .ok_or_else(|| anyhow::anyhow!("激活的模型 {id} 不在注册表中"))?;
+        if !entry.downloaded {
+            bail!("激活的模型 {id} 尚未下载");
+        }
+        Ok(self.store_dir.join(kind.dir_name()).join(&entry.info.filename))
+    }
+
+    /// 下载模型
+    pub fn download_model(&mut self, model_id: &str) -> Result<()> {
+        let entry = self
+            .registry
+            .models
+            .iter()
+            .find(|m| m.info.id == model_id)
+            .ok_or_else(|| anyhow::anyhow!("模型 {model_id} 不在清单中"))?
+            .clone();
+
+        if entry.downloaded {
+            return Ok(()); // 已下载
+        }
+
+        // 配额检查
+        if self.used_space() + entry.info.size > self.quota {
+            bail!(
+                "磁盘配额不足: 已用 {} + 需要 {} > 配额 {}",
+                self.used_space(),
+                entry.info.size,
+                self.quota
+            );
+        }
+
+        // 确保目录存在
+        let dir = self.store_dir.join(entry.info.kind.dir_name());
+        fs::create_dir_all(&dir)?;
+        let dest = dir.join(&entry.info.filename);
+
+        // 下载（简单实现，不带断点续传；生产环境应加）
+        println!("[ModelManager] 下载模型 {} from {}", model_id, entry.info.url);
+        let resp = ureq::get(&entry.info.url)
+            .timeout(std::time::Duration::from_secs(600))
+            .call()
+            .context("下载模型失败")?;
+
+        let mut file = fs::File::create(&dest)?;
+        std::io::copy(&mut resp.into_reader(), &mut file)?;
+
+        // 标记为已下载
+        if let Some(m) = self.registry.models.iter_mut().find(|m| m.info.id == model_id) {
+            m.downloaded = true;
+            m.download_progress = 100;
+        }
+        self.save_registry()?;
+        println!("[ModelManager] 模型 {} 下载完成", model_id);
+        Ok(())
+    }
+
+    /// 切换激活模型
+    pub fn switch_model(&mut self, kind: ModelKind, model_id: &str) -> Result<()> {
+        let entry = self
+            .registry
+            .models
+            .iter()
+            .find(|m| m.info.id == model_id && m.info.kind == kind)
+            .ok_or_else(|| anyhow::anyhow!("模型 {model_id} 不在 {kind:?} 清单中"))?;
+
+        if !entry.downloaded {
+            bail!("模型 {model_id} 尚未下载，无法切换");
+        }
+
+        let old = self.registry.active.insert(kind, model_id.to_string());
+        self.save_registry()?;
+        if let Some(old_id) = old {
+            println!("[ModelManager] {kind:?} 模型切换: {old_id} → {model_id}");
+        } else {
+            println!("[ModelManager] {kind:?} 模型激活: {model_id}");
+        }
+        Ok(())
+    }
+
+    /// 删除模型（当前激活的模型不可删除）
+    pub fn delete_model(&mut self, model_id: &str) -> Result<()> {
+        let entry = self
+            .registry
+            .models
+            .iter()
+            .find(|m| m.info.id == model_id)
+            .ok_or_else(|| anyhow::anyhow!("模型 {model_id} 不在清单中"))?;
+
+        // 保护当前激活模型
+        if let Some(active_id) = self.registry.active.get(&entry.info.kind) {
+            if active_id == model_id {
+                bail!("模型 {model_id} 正在使用中，无法删除（请先切换到其他模型）");
+            }
+        }
+
+        if !entry.downloaded {
+            bail!("模型 {model_id} 未下载，无需删除");
+        }
+
+        // 删除文件
+        let path = self
+            .store_dir
+            .join(entry.info.kind.dir_name())
+            .join(&entry.info.filename);
+        if path.exists() {
+            fs::remove_file(&path)?;
+        }
+
+        // 更新注册表
+        if let Some(m) = self.registry.models.iter_mut().find(|m| m.info.id == model_id) {
+            m.downloaded = false;
+            m.download_progress = 0;
+        }
+        self.save_registry()?;
+        println!("[ModelManager] 模型 {} 已删除", model_id);
+        Ok(())
+    }
+
+    /// 获取注册表（用于 HTTP API 返回）
+    pub fn registry(&self) -> &ModelRegistry {
+        &self.registry
+    }
+}
+
+/// 仓库清单索引（远程静态清单文件）
+#[derive(Debug, Clone, Deserialize)]
+pub struct RepoModelIndex {
+    pub models: Vec<ModelInfo>,
+}
+
+impl ModelManager {
+    /// 从远程仓库拉取最新清单（更新本地注册表，不触发下载）
+    pub fn fetch_repo_index(repo_url: &str) -> Result<RepoModelIndex> {
+        let resp = ureq::get(&format!("{repo_url}/models/index.json"))
+            .timeout(std::time::Duration::from_secs(30))
+            .call()
+            .context("拉取模型清单失败")?;
+        let index: RepoModelIndex = serde_json::from_reader(resp.into_reader())?;
+        Ok(index)
+    }
+}
diff --git a/src/plugins/http/mod.rs b/src/plugins/http/mod.rs
index 9aabf80..5e1da4b 100644
--- a/src/plugins/http/mod.rs
+++ b/src/plugins/http/mod.rs
@@ -48,6 +48,10 @@ pub(crate) struct HttpState {
     ws_events: broadcast::Sender<String>,
     /// 动态插件管理状态（由 Custom 消息更新）
     plugin_states: Mutex<Vec<crate::core::service_manager::PluginStateInfo>>,
+    /// AI 对话管线（HTTP 路由直接调用）
+    ai_pipeline: Mutex<Option<std::sync::Arc<crate::plugins::ai::ChatPipeline>>>,
+    /// AI 模型管理器共享句柄（HTTP 模型管理 API 调用）
+    ai_models: Mutex<Option<std::sync::Arc<std::sync::Mutex<crate::plugins::ai::model_manager::ModelManager>>>>,
 }
 
 impl HttpState {
@@ -75,9 +79,40 @@ impl HttpState {
             ble_ready: AtomicBool::new(false),
             ws_events,
             plugin_states: Mutex::new(Vec::new()),
+            ai_pipeline: Mutex::new(None),
+            ai_models: Mutex::new(None),
         }
     }
 
+    /// 注册 AI 对话管线（AiPlugin init 时调用）
+    pub(crate) fn register_ai_pipeline(&self, pipeline: std::sync::Arc<crate::plugins::ai::ChatPipeline>) {
+        if let Ok(mut slot) = self.ai_pipeline.lock() {
+            *slot = Some(pipeline);
+        }
+    }
+
+    /// 注册 AI 模型管理器（AiPlugin init 时调用）
+    pub(crate) fn register_ai_models(
+        &self,
+        models: std::sync::Arc<std::sync::Mutex<crate::plugins::ai::model_manager::ModelManager>>,
+    ) {
+        if let Ok(mut slot) = self.ai_models.lock() {
+            *slot = Some(models);
+        }
+    }
+
+    /// 获取 AI 对话管线（HTTP 路由调用）
+    pub(crate) fn ai_pipeline(&self) -> Option<std::sync::Arc<crate::plugins::ai::ChatPipeline>> {
+        self.ai_pipeline.lock().ok().and_then(|slot| slot.clone())
+    }
+
+    /// 获取 AI 模型管理器（HTTP 路由调用）
+    pub(crate) fn ai_models(
+        &self,
+    ) -> Option<std::sync::Arc<std::sync::Mutex<crate::plugins::ai::model_manager::ModelManager>>> {
+        self.ai_models.lock().ok().and_then(|slot| slot.clone())
+    }
+
     fn publish_wifi_result(&self, payload: String) {
         if let Ok(mut state) = self.wifi_response.lock() {
             state.version += 1;
@@ -210,6 +245,10 @@ pub struct HttpPlugin {
     state: Option<Arc<HttpState>>,
     shutdown_tx: Option<oneshot::Sender<()>>,
     server_thread: Option<JoinHandle<()>>,
+    /// AI 对话管线（main.rs 注册时传入，init 时存入 HttpState）
+    ai_pipeline: Option<std::sync::Arc<crate::plugins::ai::ChatPipeline>>,
+    /// AI 模型管理器（main.rs 注册时传入）
+    ai_models: Option<std::sync::Arc<std::sync::Mutex<crate::plugins::ai::model_manager::ModelManager>>>,
 }
 
 impl HttpPlugin {
@@ -219,8 +258,20 @@ impl HttpPlugin {
             state: None,
             shutdown_tx: None,
             server_thread: None,
+            ai_pipeline: None,
+            ai_models: None,
         }
     }
+
+    /// 设置 AI 对话管线（main.rs 注册 AiPlugin 后调用）
+    pub fn set_ai_pipeline(&mut self, pipeline: std::sync::Arc<crate::plugins::ai::ChatPipeline>) {
+        self.ai_pipeline = Some(pipeline);
+    }
+
+    /// 设置 AI 模型管理器（main.rs 注册 AiPlugin 后调用）
+    pub fn set_ai_models(&mut self, models: std::sync::Arc<std::sync::Mutex<crate::plugins::ai::model_manager::ModelManager>>) {
+        self.ai_models = Some(models);
+    }
 }
 
 impl Default for HttpPlugin {
@@ -248,7 +299,15 @@ impl Plugin for HttpPlugin {
     }
 
     fn init(&mut self, ctx: PluginContext) -> Result<()> {
-        self.state = Some(Arc::new(HttpState::new(Arc::clone(&ctx.config))));
+        let state = Arc::new(HttpState::new(Arc::clone(&ctx.config)));
+        // 注册 AI 句柄（如果 main.rs 已设置）
+        if let Some(pipeline) = self.ai_pipeline.take() {
+            state.register_ai_pipeline(pipeline);
+        }
+        if let Some(ai_models) = self.ai_models.take() {
+            state.register_ai_models(ai_models);
+        }
+        self.state = Some(state);
         self.ctx = Some(ctx);
         Ok(())
     }
diff --git a/src/plugins/http/routes.rs b/src/plugins/http/routes.rs
index 9c94430..de71429 100644
--- a/src/plugins/http/routes.rs
+++ b/src/plugins/http/routes.rs
@@ -1,7 +1,7 @@
 use super::HttpState;
 use crate::core::config::{self, AppConfig};
 use crate::core::dispatch;
-use crate::core::message::{Destination, Envelope, Message, PlayerCommand, WifiCommand};
+use crate::core::message::{ChatInput, ChatRequest, Destination, Envelope, Message, PlayerCommand, WifiCommand};
 use bytes::Buf;
 use futures_util::{SinkExt, StreamExt, TryStreamExt};
 use serde::de::DeserializeOwned;
@@ -145,7 +145,15 @@ pub(crate) fn build_routes(
         .or(file_mkdir_route(Arc::clone(&state)))
         .boxed();
 
-    let api = core_api.or(media_api).or(plugin_api).or(file_api);
+    let ai_api = chat_text_route(Arc::clone(&state))
+        .or(chat_audio_route(Arc::clone(&state)))
+        .or(models_list_route(Arc::clone(&state)))
+        .or(model_download_route(Arc::clone(&state)))
+        .or(model_switch_route(Arc::clone(&state)))
+        .or(model_delete_route(Arc::clone(&state)))
+        .boxed();
+
+    let api = core_api.or(media_api).or(plugin_api).or(file_api).or(ai_api);
 
     root_route()
         .or(download_route(Arc::clone(&state)))
@@ -1954,6 +1962,270 @@ async fn send_plugin_command(
     }
 }
 
+// ── AI 对话 API (M2.1) ──
+
+#[derive(Deserialize)]
+struct ChatTextRequest {
+    #[serde(default)]
+    session_id: Option<String>,
+    text: String,
+}
+
+/// POST /api/chat/text — 文字对话（Web 端主路径）
+fn chat_text_route(
+    state: Arc<HttpState>,
+) -> impl Filter<Extract = impl Reply, Error = warp::Rejection> + Clone {
+    warp::path!("api" / "chat" / "text")
+        .and(warp::post())
+        .and(warp::body::json::<ChatTextRequest>())
+        .and(with_state(state))
+        .and_then(|req: ChatTextRequest, state: Arc<HttpState>| async move {
+            let pipeline = match state.ai_pipeline() {
+                Some(p) => p,
+                None => {
+                    return Ok::<_, Infallible>(error_json(
+                        StatusCode::SERVICE_UNAVAILABLE,
+                        "AI 插件未就绪",
+                    ))
+                }
+            };
+
+            let config = state.config();
+            let session_id = req
+                .session_id
+                .unwrap_or_else(|| format!("web_{}", std::process::id()));
+            let chat_req = ChatRequest {
+                session_id,
+                input: ChatInput::Text {
+                    content: req.text,
+                },
+                persona_prompt: config.character.persona_prompt.clone(),
+                max_tokens: config.character.max_tokens,
+            };
+
+            // AI 管线是同步阻塞调用（ASR/LLM/TTS 子进程），用 spawn_blocking 避免阻塞 tokio reactor
+            let resp = tokio::task::spawn_blocking(move || pipeline.run(&chat_req))
+                .await
+                .unwrap_or_else(|e| ChatResponse {
+                    session_id: String::new(),
+                    transcription: None,
+                    reply_text: String::new(),
+                    reply_audio_path: None,
+                    error: Some(format!("AI 管线执行失败: {e}")),
+                });
+
+            if resp.error.is_some() {
+                Ok(json_response(StatusCode::INTERNAL_SERVER_ERROR, &resp))
+            } else {
+                Ok(json_response(StatusCode::OK, &resp))
+            }
+        })
+}
+
+/// POST /api/chat/audio — 语音对话（App 主路径，上传 wav）
+fn chat_audio_route(
+    state: Arc<HttpState>,
+) -> impl Filter<Extract = impl Reply, Error = warp::Rejection> + Clone {
+    warp::path!("api" / "chat" / "audio")
+        .and(warp::post())
+        .and(warp::body::content_length_limit(20 * 1024 * 1024))
+        .and(warp::body::bytes())
+        .and(with_state(state))
+        .and_then(|bytes: bytes::Bytes, state: Arc<HttpState>| async move {
+            let pipeline = match state.ai_pipeline() {
+                Some(p) => p,
+                None => {
+                    return Ok::<_, Infallible>(error_json(
+                        StatusCode::SERVICE_UNAVAILABLE,
+                        "AI 插件未就绪",
+                    ))
+                }
+            };
+
+            let config = state.config();
+            let tmp_dir = config.ai.tmp_dir.clone();
+            let _ = std::fs::create_dir_all(&tmp_dir);
+            let audio_path = tmp_dir.join(format!(
+                "asr_{}.wav",
+                std::time::SystemTime::now()
+                    .duration_since(std::time::UNIX_EPOCH)
+                    .map(|d| d.as_millis())
+                    .unwrap_or(0)
+            ));
+            if let Err(e) = std::fs::write(&audio_path, &bytes) {
+                return Ok(error_json(
+                    StatusCode::INTERNAL_SERVER_ERROR,
+                    &format!("保存音频失败: {e}"),
+                ));
+            }
+
+            let session_id = format!("app_{}", std::process::id());
+            let chat_req = ChatRequest {
+                session_id,
+                input: ChatInput::Audio {
+                    path: audio_path.to_string_lossy().into_owned(),
+                },
+                persona_prompt: config.character.persona_prompt.clone(),
+                max_tokens: config.character.max_tokens,
+            };
+
+            let audio_path_clone = audio_path.clone();
+            let resp = tokio::task::spawn_blocking(move || pipeline.run(&chat_req))
+                .await
+                .unwrap_or_else(|e| ChatResponse {
+                    session_id: String::new(),
+                    transcription: None,
+                    reply_text: String::new(),
+                    reply_audio_path: None,
+                    error: Some(format!("AI 管线执行失败: {e}")),
+                });
+
+            // 清理上传的临时音频
+            let _ = std::fs::remove_file(&audio_path_clone);
+
+            if resp.error.is_some() {
+                Ok(json_response(StatusCode::INTERNAL_SERVER_ERROR, &resp))
+            } else {
+                Ok(json_response(StatusCode::OK, &resp))
+            }
+        })
+}
+
+// ── AI 模型管理 API (M2.1) ──
+
+/// GET /api/models — 列出所有模型
+fn models_list_route(
+    state: Arc<HttpState>,
+) -> impl Filter<Extract = impl Reply, Error = warp::Rejection> + Clone {
+    warp::path!("api" / "models")
+        .and(warp::get())
+        .and(with_state(state))
+        .and_then(|state: Arc<HttpState>| async move {
+            let models = match state.ai_models() {
+                Some(m) => m,
+                None => {
+                    return Ok::<_, Infallible>(error_json(
+                        StatusCode::SERVICE_UNAVAILABLE,
+                        "AI 插件未就绪",
+                    ))
+                }
+            };
+            let mgr = models.lock().unwrap();
+            let registry = mgr.registry();
+            let used = mgr.used_space();
+            let quota = mgr.quota();
+            let result = serde_json::json!({
+                "models": registry.models,
+                "active": registry.active,
+                "used_space": used,
+                "quota": quota,
+            });
+            Ok(json_response(StatusCode::OK, &result))
+        })
+}
+
+#[derive(Deserialize)]
+struct ModelActionRequest {
+    model_id: String,
+}
+
+/// POST /api/models/download — 下载模型
+fn model_download_route(
+    state: Arc<HttpState>,
+) -> impl Filter<Extract = impl Reply, Error = warp::Rejection> + Clone {
+    warp::path!("api" / "models" / "download")
+        .and(warp::post())
+        .and(warp::body::json::<ModelActionRequest>())
+        .and(with_state(state))
+        .and_then(|req: ModelActionRequest, state: Arc<HttpState>| async move {
+            let models = match state.ai_models() {
+                Some(m) => m,
+                None => {
+                    return Ok::<_, Infallible>(error_json(
+                        StatusCode::SERVICE_UNAVAILABLE,
+                        "AI 插件未就绪",
+                    ))
+                }
+            };
+            // 下载在独立线程执行，避免阻塞 HTTP
+            let models_clone = models.clone();
+            let model_id = req.model_id.clone();
+            std::thread::spawn(move || {
+                let mut mgr = models_clone.lock().unwrap();
+                if let Err(e) = mgr.download_model(&model_id) {
+                    eprintln!("[HttpPlugin] 模型下载失败 {model_id}: {e}");
+                }
+            });
+            Ok::<_, Infallible>(success_json(format!("模型 {} 下载已启动", req.model_id)))
+        })
+}
+
+/// POST /api/models/switch — 切换激活模型
+fn model_switch_route(
+    state: Arc<HttpState>,
+) -> impl Filter<Extract = impl Reply, Error = warp::Rejection> + Clone {
+    warp::path!("api" / "models" / "switch")
+        .and(warp::post())
+        .and(warp::body::json::<serde_json::Value>())
+        .and(with_state(state))
+        .and_then(|body: serde_json::Value, state: Arc<HttpState>| async move {
+            let models = match state.ai_models() {
+                Some(m) => m,
+                None => {
+                    return Ok::<_, Infallible>(error_json(
+                        StatusCode::SERVICE_UNAVAILABLE,
+                        "AI 插件未就绪",
+                    ))
+                }
+            };
+            let model_id = match body.get("model_id").and_then(|v| v.as_str()) {
+                Some(s) => s.to_string(),
+                None => return Ok(error_json(StatusCode::BAD_REQUEST, "缺少 model_id")),
+            };
+            let kind_str = match body.get("kind").and_then(|v| v.as_str()) {
+                Some(s) => s,
+                None => return Ok(error_json(StatusCode::BAD_REQUEST, "缺少 kind")),
+            };
+            let kind = match kind_str {
+                "llm" => crate::plugins::ai::model_manager::ModelKind::Llm,
+                "asr" => crate::plugins::ai::model_manager::ModelKind::Asr,
+                "tts" => crate::plugins::ai::model_manager::ModelKind::Tts,
+                _ => return Ok(error_json(StatusCode::BAD_REQUEST, "kind 必须为 llm/asr/tts")),
+            };
+            let mut mgr = models.lock().unwrap();
+            match mgr.switch_model(kind, &model_id) {
+                Ok(()) => Ok(success_json(format!("模型已切换为 {}", model_id))),
+                Err(e) => Ok(error_json(StatusCode::BAD_REQUEST, &e.to_string())),
+            }
+        })
+}
+
+/// POST /api/models/delete — 删除模型
+fn model_delete_route(
+    state: Arc<HttpState>,
+) -> impl Filter<Extract = impl Reply, Error = warp::Rejection> + Clone {
+    warp::path!("api" / "models" / "delete")
+        .and(warp::post())
+        .and(warp::body::json::<ModelActionRequest>())
+        .and(with_state(state))
+        .and_then(|req: ModelActionRequest, state: Arc<HttpState>| async move {
+            let models = match state.ai_models() {
+                Some(m) => m,
+                None => {
+                    return Ok::<_, Infallible>(error_json(
+                        StatusCode::SERVICE_UNAVAILABLE,
+                        "AI 插件未就绪",
+                    ))
+                }
+            };
+            let mut mgr = models.lock().unwrap();
+            match mgr.delete_model(&req.model_id) {
+                Ok(()) => Ok(success_json(format!("模型 {} 已删除", req.model_id))),
+                Err(e) => Ok(error_json(StatusCode::BAD_REQUEST, &e.to_string())),
+            }
+        })
+}
+
 fn json_response<T: Serialize>(status: StatusCode, payload: &T) -> warp::reply::Response {
     warp::reply::with_status(warp::reply::json(payload), status).into_response()
 }
diff --git a/src/plugins/mod.rs b/src/plugins/mod.rs
index da97f0b..c87a006 100644
--- a/src/plugins/mod.rs
+++ b/src/plugins/mod.rs
@@ -1,3 +1,4 @@
+pub mod ai;
 pub mod ble;
 pub mod device;
 pub mod http;