```

feat(cli): add direct model inference via new run subcommand

- Add `aha run` CLI subcommand for direct model inference without HTTP service
- Support multiple models including Qwen series, OCR models, ASR models, and voice generation
- Implement input/output handling with file path support and auto-generation
- Add comprehensive documentation in CLI_USAGE.md with examples
- Include performance timing for model loading and inference operations
- Add macOS build target to Makefile with Metal support
```

CHANGELOG.md

@@ -1,18 +1,17 @@
 # Changelog
 
-All notable changes to this project will be documented in this file.
-
-The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
-and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
 ### Added
 
-- **CLI Subcommand Support**: Added three new subcommands for better command organization:
-  - `aha cli` - Download model and start HTTP service (default, backward compatible)
-  - `aha serv` - Start HTTP service only (requires `--weight-path`)
-  - `aha download` - Download model only (no service start)
+- **CLI `run` Subcommand**: Direct model inference from CLI without HTTP service overhead:
+  - `aha run` - Run model inference directly
+  - `-m, --model <MODEL>` - Specify which model to use
+  - `-in, --input <INPUT>` - Input text or file path (model-specific interpretation)
+  - `-out, --output <OUTPUT>` - Output file path (optional, auto-generated if not specified)
+  - `--weight-path <WEIGHT_PATH>` - Local model weight path (required)
+
 
 ### Changed
 

Makefile

@@ -10,6 +10,10 @@ build:
 	@echo "Building project..."
 	@cargo build
 
+build_mac:
+	@echo "Building project for macOS..."
+	@cargo build --features metal  --release
+
 test:
 	@echo "Running tests..."
 	@cargo test

docs/CLI_USAGE.md

@@ -59,6 +59,46 @@ aha cli -m qwen3vl-2b --weight-path /path/to/model
 aha -m qwen3vl-2b
 ```
 
+### run - 直接模型推理
+
+直接运行模型推理，无需启动 HTTP 服务。适用于一次性推理任务或批处理。
+
+**语法：**
+```bash
+aha run [OPTIONS] --model <MODEL> --input <INPUT> --weight-path <WEIGHT_PATH>
+```
+
+**选项：**
+
+| 选项 | 说明 | 默认值 |
+|------|------|--------|
+| `-m, --model <MODEL>` | 模型类型（必选） | - |
+| `-in, --input <INPUT>` | 输入文本或文件路径（模型特定解释） | - |
+| `-out, --output <OUTPUT>` | 输出文件路径（可选，未指定则自动生成） | - |
+| `--weight-path <WEIGHT_PATH>` | 本地模型权重路径（必选） | - |
+
+**示例：**
+
+```bash
+# VoxCPM1.5 文字转语音
+aha run -m voxcpm1.5 -in "太阳当空照" -out output.wav --weight-path /path/to/model
+
+# VoxCPM1.5 从文件读取输入
+aha run -m voxcpm1.5 -in "file://./input.txt" --weight-path /path/to/model
+
+# MiniCPM4 文本生成
+aha run -m minicpm4-0.5b -in "你好" --weight-path /path/to/model
+
+# DeepSeek OCR 图片识别
+aha run -m deepseek-ocr -in "image.jpg" --weight-path /path/to/model
+
+# RMBG2.0 背景移除
+aha run -m RMBG2.0 -in "photo.png" -out "no_bg.png" --weight-path /path/to/model
+
+# GLM-ASR 语音识别
+aha run -m glm-asr-nano-2512 -in "audio.wav" -in "请转写这段音频" --weight-path /path/to/model
+```
+
 ### serv - 启动服务
 
 仅启动 HTTP 服务，不下载模型。必须通过 `--weight-path` 指定本地模型路径。

src/exec/deepseek_ocr.rs

@@ -0,0 +1,58 @@
+//! DeepSeek-OCR exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, deepseek_ocr::generate::DeepseekOCRGenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct DeepSeekORExec;
+
+impl ExecModel for DeepSeekORExec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let input_path = if input.starts_with("file://") {
+            input.to_string()
+        } else {
+            format!("file://{}", input)
+        };
+
+        let i_start = Instant::now();
+        let mut model = DeepseekOCRGenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let message = format!(
+            r#"{{
+            "model": "deepseek-ocr",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": [
+                        {{
+                            "type": "image",
+                            "image_url": {{
+                                "url": "{}"
+                            }}
+                        }}
+                    ]
+                }}
+            ]
+        }}"#,
+            input_path
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let result = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/fun_asr_nano.rs

@@ -0,0 +1,70 @@
+//! Fun-ASR-Nano-2512 exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, fun_asr_nano::generate::FunAsrNanoGenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct FunASRNanoExec;
+
+impl ExecModel for FunASRNanoExec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let target_text = if input.starts_with("file://") {
+            let path = &input[7..];
+            std::fs::read_to_string(path)?
+        } else {
+            input.to_string()
+        };
+
+        let i_start = Instant::now();
+        let mut model = FunAsrNanoGenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        // Create ChatCompletionParameters for ASR
+        let input_url = if input.starts_with("http://") || input.starts_with("https://") || input.starts_with("file://") {
+            input.to_string()
+        } else {
+            format!("file://{}", input)
+        };
+
+        let message = format!(
+            r#"{{
+            "model": "fun-asr-nano",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": [
+                        {{
+                            "type": "audio",
+                            "audio_url": {{
+                                "url": "{}"
+                            }}
+                        }},
+                        {{
+                            "type": "text",
+                            "text": "{}"
+                        }}
+                    ]
+                }}
+            ]
+        }}"#,
+            input_url, target_text
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let res = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", res);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", res))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/glm_asr_nano.rs

@@ -0,0 +1,71 @@
+//! GLM-ASR-Nano-2512 exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, glm_asr_nano::generate::GlmAsrNanoGenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct GlmASRNanoExec;
+
+impl ExecModel for GlmASRNanoExec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let target_text = if input.starts_with("file://") {
+            let path = &input[7..];
+            std::fs::read_to_string(path)?
+        } else {
+            input.to_string()
+        };
+
+        let i_start = Instant::now();
+        let mut model = GlmAsrNanoGenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        // Create ChatCompletionParameters for ASR
+        // Input should be an audio file path
+        let input_url = if input.starts_with("http://") || input.starts_with("https://") || input.starts_with("file://") {
+            input.to_string()
+        } else {
+            format!("file://{}", input)
+        };
+
+        let message = format!(
+            r#"{{
+            "model": "glm-asr-nano",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": [
+                        {{
+                            "type": "audio",
+                            "audio_url": {{
+                                "url": "{}"
+                            }}
+                        }},
+                        {{
+                            "type": "text",
+                            "text": "{}"
+                        }}
+                    ]
+                }}
+            ]
+        }}"#,
+            input_url, target_text
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let res = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", res);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", res))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/hunyuan_ocr.rs

@@ -0,0 +1,58 @@
+//! Hunyuan-OCR exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, hunyuan_ocr::generate::HunyuanOCRGenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct HunyuanORExec;
+
+impl ExecModel for HunyuanORExec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let input_path = if input.starts_with("file://") {
+            input.to_string()
+        } else {
+            format!("file://{}", input)
+        };
+
+        let i_start = Instant::now();
+        let mut model = HunyuanOCRGenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let message = format!(
+            r#"{{
+            "model": "hunyuan-ocr",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": [
+                        {{
+                            "type": "image",
+                            "image_url": {{
+                                "url": "{}"
+                            }}
+                        }}
+                    ]
+                }}
+            ]
+        }}"#,
+            input_path
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let result = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/minicpm4.rs

@@ -0,0 +1,55 @@
+//! MiniCPM4-0.5B exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, minicpm4::generate::MiniCPMGenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct MiniCPM4Exec;
+
+impl ExecModel for MiniCPM4Exec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let target_text = if input.starts_with("file://") {
+            let path = &input[7..];
+            std::fs::read_to_string(path)?
+        } else {
+            input.to_string()
+        };
+
+        let i_start = Instant::now();
+        let mut model = MiniCPMGenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let message = format!(
+            r#"{{
+            "temperature": 0.3,
+            "top_p": 0.8,
+            "model": "minicpm4",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": "{}"
+                }}
+            ]
+        }}"#,
+            target_text.replace('"', "\\\"")
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let result = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        // Print result
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/mod.rs

@@ -0,0 +1,37 @@
+//! CLI exec module for direct model inference
+//!
+//! This module provides model-specific exec implementations for the `run` subcommand.
+//! Each model has its own exec module that handles input/output parsing and model invocation.
+
+pub mod deepseek_ocr;
+pub mod fun_asr_nano;
+pub mod glm_asr_nano;
+pub mod hunyuan_ocr;
+pub mod minicpm4;
+pub mod paddleocr_vl;
+pub mod qwen2_5vl;
+pub mod qwen3;
+pub mod qwen3vl;
+pub mod rmbg2_0;
+pub mod voxcpm;
+pub mod voxcpm1_5;
+
+use anyhow::Result;
+
+/// Trait for model exec implementations
+///
+/// Each model exec module implements this trait to provide
+/// model-specific inference logic for CLI `run` commands.
+pub trait ExecModel {
+    /// Run inference with the given input and output parameters
+    ///
+    /// # Arguments
+    /// * `input` - Input text or file path (interpretation is model-specific)
+    /// * `output` - Optional output file path (if None, model will auto-generate)
+    /// * `weight_path` - Path to the model weights
+    ///
+    /// # Returns
+    /// * `Ok(())` on success
+    /// * `Err(anyhow::Error)` on failure
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()>;
+}

src/exec/paddleocr_vl.rs

@@ -0,0 +1,58 @@
+//! PaddleOCR-VL exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, paddleocr_vl::generate::PaddleOCRVLGenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct PaddleOVLExec;
+
+impl ExecModel for PaddleOVLExec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let input_path = if input.starts_with("file://") {
+            input.to_string()
+        } else {
+            format!("file://{}", input)
+        };
+
+        let i_start = Instant::now();
+        let mut model = PaddleOCRVLGenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let message = format!(
+            r#"{{
+            "model": "paddleocr-vl",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": [
+                        {{
+                            "type": "image",
+                            "image_url": {{
+                                "url": "{}"
+                            }}
+                        }}
+                    ]
+                }}
+            ]
+        }}"#,
+            input_path
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let result = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/qwen2_5vl.rs

@@ -0,0 +1,52 @@
+//! Qwen2.5VL-3B exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, qwen2_5vl::generate::Qwen2_5VLGenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct Qwen2_5vlExec;
+
+impl ExecModel for Qwen2_5vlExec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let target_text = if input.starts_with("file://") {
+            let path = &input[7..];
+            std::fs::read_to_string(path)?
+        } else {
+            input.to_string()
+        };
+
+        let i_start = Instant::now();
+        let mut model = Qwen2_5VLGenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let message = format!(
+            r#"{{
+            "model": "qwen2.5vl",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": "{}"
+                }}
+            ]
+        }}"#,
+            target_text.replace('"', "\\\"")
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let result = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/qwen3.rs

@@ -0,0 +1,52 @@
+//! Qwen3-0.6B exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, qwen3::generate::Qwen3GenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct Qwen3Exec;
+
+impl ExecModel for Qwen3Exec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let target_text = if input.starts_with("file://") {
+            let path = &input[7..];
+            std::fs::read_to_string(path)?
+        } else {
+            input.to_string()
+        };
+
+        let i_start = Instant::now();
+        let mut model = Qwen3GenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let message = format!(
+            r#"{{
+            "model": "qwen3",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": "{}"
+                }}
+            ]
+        }}"#,
+            target_text.replace('"', "\\\"")
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let result = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/qwen3vl.rs

@@ -0,0 +1,52 @@
+//! Qwen3VL-2B exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::{GenerateModel, qwen3vl::generate::Qwen3VLGenerateModel};
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct Qwen3vlExec;
+
+impl ExecModel for Qwen3vlExec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let target_text = if input.starts_with("file://") {
+            let path = &input[7..];
+            std::fs::read_to_string(path)?
+        } else {
+            input.to_string()
+        };
+
+        let i_start = Instant::now();
+        let mut model = Qwen3VLGenerateModel::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let message = format!(
+            r#"{{
+            "model": "qwen3vl",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": "{}"
+                }}
+            ]
+        }}"#,
+            target_text.replace('"', "\\\"")
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let result = model.generate(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        println!("Result: {:?}", result);
+
+        if let Some(out) = output {
+            std::fs::write(out, format!("{:?}", result))?;
+            println!("Output saved to: {}", out);
+        }
+
+        Ok(())
+    }
+}

src/exec/rmbg2_0.rs

@@ -0,0 +1,72 @@
+//! RMBG2.0 exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::rmbg2_0::generate::RMBG2_0Model;
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct RMBG2_0Exec;
+
+impl ExecModel for RMBG2_0Exec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let input_path = if input.starts_with("file://") {
+            input.to_string()
+        } else {
+            format!("file://{}", input)
+        };
+
+        let i_start = Instant::now();
+        let model = RMBG2_0Model::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        // Create ChatCompletionParameters for image background removal
+        let message = format!(
+            r#"{{
+            "model": "rmbg2.0",
+            "messages": [
+                {{
+                    "role": "user",
+                    "content": [
+                        {{
+                            "type": "image",
+                            "image_url": {{
+                                "url": "{}"
+                            }}
+                        }}
+                    ]
+                }}
+            ]
+        }}"#,
+            input_path
+        );
+        let mes = serde_json::from_str(&message)?;
+
+        let i_start = Instant::now();
+        let result = model.inference(mes)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        let output_path = if let Some(out) = output {
+            out.to_string()
+        } else {
+            let timestamp = std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)?
+                .as_secs();
+            format!("rmbg_{}.png", timestamp)
+        };
+
+        // Save all result images
+        for (i, img) in result.iter().enumerate() {
+            let path = if result.len() == 1 {
+                output_path.clone()
+            } else {
+                format!("{}_{}.png", output_path.trim_end_matches(".png"), i)
+            };
+            img.save(&path)?;
+            println!("Output saved to: {}", path);
+        }
+
+        Ok(())
+    }
+}

src/exec/voxcpm.rs

@@ -0,0 +1,54 @@
+//! VoxCPM exec implementation for CLI `run` subcommand
+
+use crate::exec::ExecModel;
+use crate::models::voxcpm::generate::VoxCPMGenerate;
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct VoxCPMExec;
+
+impl ExecModel for VoxCPMExec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let target_text = if input.starts_with("file://") {
+            let path = &input[7..];
+            std::fs::read_to_string(path)?
+        } else {
+            input.to_string()
+        };
+
+        let i_start = Instant::now();
+        let mut voxcpm_generate = VoxCPMGenerate::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let i_start = Instant::now();
+        let audio = voxcpm_generate.inference(
+            target_text,
+            Some("啥子小师叔，打狗还要看主人，你再要继续，我就是你的对手".to_string()), //todo args
+            Some("file://./assets/audio/voice_01.wav".to_string()),  //todo args
+            2,
+            100,   // max_len (voxcpm uses 100 vs voxcpm1.5's 4096)
+            10,
+            2.0,
+            6.0,
+        )?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        let output_path = if let Some(out) = output {
+            out.to_string()
+        } else {
+            let timestamp = std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)?
+                .as_secs();
+            format!("voxcpm_{}.wav", timestamp)
+        };
+
+        let sample_rate = voxcpm_generate.sample_rate();
+        crate::utils::audio_utils::save_wav(&audio, &output_path, sample_rate as u32)?;
+
+        println!("Output saved to: {}", output_path);
+
+        Ok(())
+    }
+}

src/exec/voxcpm1_5.rs

@@ -0,0 +1,59 @@
+//! VoxCPM1.5 exec implementation for CLI `run` subcommand
+//!
+//! This module handles VoxCPM1.5 model inference for direct CLI execution.
+//! Input/output parameter interpretation is handled here as per the design:
+//! - Input can be text content or a file path (with `file://` prefix)
+//! - Output can be a file path or will be auto-generated if not specified
+
+use crate::exec::ExecModel;
+use crate::models::voxcpm::generate::VoxCPMGenerate;
+use anyhow::{Ok, Result};
+use std::time::Instant;
+
+pub struct VoxCPM1_5Exec;
+
+impl ExecModel for VoxCPM1_5Exec {
+    fn run(input: &str, output: Option<&str>, weight_path: &str) -> Result<()> {
+        let target_text = if input.starts_with("file://") {
+            let path = &input[7..];
+            std::fs::read_to_string(path)?
+        } else {
+            input.to_string()
+        };
+
+        let i_start = Instant::now();
+        let mut voxcpm_generate = VoxCPMGenerate::init(weight_path, None, None)?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in load model is: {:?}", i_duration);
+
+        let i_start = Instant::now();
+        let audio = voxcpm_generate.inference(
+            target_text,
+            Some("啥子小师叔，打狗还要看主人，你再要继续，我就是你的对手".to_string()), //todo args
+            Some("file://./assets/audio/voice_01.wav".to_string()),  //todo args
+            2,
+            4096,
+            10,
+            2.0,
+            6.0,
+        )?;
+        let i_duration = i_start.elapsed();
+        println!("Time elapsed in generate is: {:?}", i_duration);
+
+        let output_path = if let Some(out) = output {
+            out.to_string()
+        } else {
+            let timestamp = std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)?
+                .as_secs();
+            format!("voxcpm1_5_{}.wav", timestamp)
+        };
+
+        let sample_rate = voxcpm_generate.sample_rate();
+        crate::utils::audio_utils::save_wav(&audio, &output_path, sample_rate as u32)?;
+
+        println!("Output saved to: {}", output_path);
+
+        Ok(())
+    }
+}

src/lib.rs

@@ -1,4 +1,5 @@
 pub mod chat_template;
+pub mod exec;
 pub mod models;
 pub mod position_embed;
 pub mod tokenizer;

src/main.rs

@@ -53,6 +53,8 @@ enum Commands {
     Serv(ServArgs),
     /// Download model only
     Download(DownloadArgs),
+    /// Run model inference directly
+    Run(RunArgs),
 }
 
 /// Common/shared arguments for server operations
@@ -117,6 +119,26 @@ struct DownloadArgs {
     download_retries: Option<u32>,
 }
 
+/// Arguments for the 'run' subcommand (direct inference)
+#[derive(Args, Debug)]
+struct RunArgs {
+    /// Model type (required)
+    #[arg(short, long)]
+    model: WhichModel,
+
+    /// Input text or file path
+    #[arg(short, long)]
+    input: String,
+
+    /// Output file path (optional)
+    #[arg(short, long)]
+    output: Option<String>,
+
+    /// Local model weight path (required)
+    #[arg(long, required = true)]
+    weight_path: String,
+}
+
 async fn download_model(model_id: &str, save_dir: &str, max_retries: u32) -> anyhow::Result<()> {
     let mut attempts = 0u32;
     loop {
@@ -222,6 +244,82 @@ async fn run_download(args: DownloadArgs) -> anyhow::Result<()> {
     Ok(())
 }
 
+/// Run the 'run' subcommand: direct model inference
+fn run_run(args: RunArgs) -> anyhow::Result<()> {
+    use aha::exec::ExecModel;
+
+    let RunArgs { model, input, output, weight_path } = args;
+
+    match model {
+        WhichModel::MiniCPM4_0_5B => {
+            use aha::exec::minicpm4::MiniCPM4Exec;
+            MiniCPM4Exec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::Qwen2_5vl3B => {
+            use aha::exec::qwen2_5vl::Qwen2_5vlExec;
+            Qwen2_5vlExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::Qwen2_5vl7B => {
+            use aha::exec::qwen2_5vl::Qwen2_5vlExec;
+            Qwen2_5vlExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::Qwen3_0_6B => {
+            use aha::exec::qwen3::Qwen3Exec;
+            Qwen3Exec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::Qwen3vl2B => {
+            use aha::exec::qwen3vl::Qwen3vlExec;
+            Qwen3vlExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::Qwen3vl4B => {
+            use aha::exec::qwen3vl::Qwen3vlExec;
+            Qwen3vlExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::Qwen3vl8B => {
+            use aha::exec::qwen3vl::Qwen3vlExec;
+            Qwen3vlExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::Qwen3vl32B => {
+            use aha::exec::qwen3vl::Qwen3vlExec;
+            Qwen3vlExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::DeepSeekOCR => {
+            use aha::exec::deepseek_ocr::DeepSeekORExec;
+            DeepSeekORExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::HunyuanOCR => {
+            use aha::exec::hunyuan_ocr::HunyuanORExec;
+            HunyuanORExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::PaddleOCRVL => {
+            use aha::exec::paddleocr_vl::PaddleOVLExec;
+            PaddleOVLExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::RMBG2_0 => {
+            use aha::exec::rmbg2_0::RMBG2_0Exec;
+            RMBG2_0Exec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::VoxCPM => {
+            use aha::exec::voxcpm::VoxCPMExec;
+            VoxCPMExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::VoxCPM1_5 => {
+            use aha::exec::voxcpm1_5::VoxCPM1_5Exec;
+            VoxCPM1_5Exec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::GlmASRNano2512 => {
+            use aha::exec::glm_asr_nano::GlmASRNanoExec;
+            GlmASRNanoExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+        WhichModel::FunASRNano2512 => {
+            use aha::exec::fun_asr_nano::FunASRNanoExec;
+            FunASRNanoExec::run(&input, output.as_deref(), &weight_path)?;
+        }
+    }
+
+    Ok(())
+}
+
 #[tokio::main]
 async fn main() -> anyhow::Result<()> {
     let cli = Cli::parse();
@@ -230,6 +328,7 @@ async fn main() -> anyhow::Result<()> {
         Some(Commands::Cli(args)) => run_cli(args).await,
         Some(Commands::Serv(args)) => run_serv(args).await,
         Some(Commands::Download(args)) => run_download(args).await,
+        Some(Commands::Run(args)) => run_run(args),
         None => {
             // Backward compatibility: when no subcommand is provided, use 'cli' behavior
             let model = cli.model.expect("Model is required (use -m or --model)");