MiMo - Xiaomi AI Model Provider for openvela/NuttX

A standalone Xiaomi MiMo large language model provider for embedded systems, built on the openvela/NuttX platform with full HTTPS + TLS support for on-device cloud inference.

English | 中文

Features

• OpenAI-compatible Chat Completions API
• Thinking mode (reasoning_content) for chain-of-thought reasoning
• Function calling / tool use support
• Interactive REPL and single-shot CLI modes
• Secure HTTPS communication via mbedTLS
• Automatic conversation history management with multi-turn support

Model

mimo-v2-flash: 309B total / 15B active parameters (MoE architecture), 262K context window, hybrid attention.

• Endpoint: https://api.xiaomimimo.com/v1/chat/completions
• Auth: Bearer <API_KEY>
• Get your key at: https://platform.xiaomimimo.com/

Project Structure

packages/demos/mimo/
├── mimo.h             # Core data structures and API definitions
├── mimo_main.c        # Entry point, CLI interaction logic
├── mimo_agent.c       # Agent layer: conversation history & tool dispatch
├── mimo_provider.c    # Provider layer: HTTP/TLS communication & JSON handling
├── Kconfig            # NuttX menu configuration
├── CMakeLists.txt     # CMake build script
├── Makefile           # Make build script
└── Make.defs          # NuttX application registration

Build & Run (Emulator Example)

The following example uses the openvela goldfish-armeabi-v7a-ap emulator target.

1. Configure

Enable DEMOS_MIMO in menuconfig and set your API key:

./build.sh vendor/openvela/boards/vela/configs/goldfish-armeabi-v7a-ap/ menuconfig
# Navigate to: Application Configuration → Demos → Xiaomi MiMo AI Provider
# Enable and set API key

After saving menuconfig, the following configs will be added to vendor/openvela/boards/vela/configs/goldfish-armeabi-v7a-ap/defconfig:

CONFIG_DEMOS_MIMO=y
CONFIG_MIMO_API_KEY="your_api_key_here"

Configuration Options

Option	Default	Description
MIMO_API_KEY	""	API key from platform.xiaomimimo.com
MIMO_MODEL	mimo-v2-flash	Model identifier
MIMO_API_BASE_URL	https://api.xiaomimimo.com/v1/chat/completions	API endpoint
MIMO_TEMPERATURE	80	Temperature x100 (80 = 0.8)
MIMO_ENABLE_THINKING	y	Enable chain-of-thought reasoning mode
MIMO_ENABLE_TOOLS	y	Enable function calling
MIMO_STACKSIZE	32768	Task stack size (32KB recommended for TLS)

2. Build

./build.sh vendor/openvela/boards/vela/configs/goldfish-armeabi-v7a-ap -j8

3. Start Emulator

./emulator.sh vela -no-windows

4. Run in NSH

# Bring up network interface
openvela-ap> ifup eth0
ifup eth0...OK

# Obtain IP address via DHCP
openvela-ap> renew eth0

# Launch MiMo (interactive mode)
openvela-ap> mimo

# Or single-shot mode
openvela-ap> mimo What is the capital of France?

Example Session

openvela-ap> mimo
[mimo] Provider initialized
[mimo] Endpoint: https://api.xiaomimimo.com/v1/chat/completions
[mimo] Thinking mode: enabled
[mimo:agent] Initialized (max_turns=10, max_history=32)

  MiMo - Xiaomi AI Assistant
  Model: mimo-v2-flash
  Thinking: on
  Type 'quit' to exit.

You> who are you
[mimo] POST https://api.xiaomimimo.com/v1/chat/completions (2318 bytes)
[mimo] TLS connecting to api.xiaomimimo.com:443
[mimo] Connecting to api.xiaomimimo.com (120.133.85.111:443)
[mimo] TCP connected, starting TLS handshake
[mimo] TLS handshake complete
[mimo] HTTP 200, response 1815 bytes
[mimo] Thinking: 好的，用户现在问"who are you"，这是英语版本的"你是谁"...
MiMo> I am MiMo, an AI assistant developed by the Xiaomi LLM Core Team.
      I'm here to help with information queries, problem-solving,
      content creation, logical reasoning, and more.
      How can I assist you today? 😊

You>

Type quit or exit to leave interactive mode.

Architecture

User Input → mimo_main → mimo_agent_chat → mimo_provider_chat → MiMo API
                              ↑                    ↓
                      Conversation History    HTTP/TLS Communication
                      Tool Dispatch Loop      JSON Request Building
                        (ReAct Loop)          Response Parsing

• Provider Layer (mimo_provider.c): Handles HTTP POST requests, TLS handshake, JSON request body construction, and response parsing. Uses NuttX webclient + mbedTLS for HTTPS.
• Agent Layer (mimo_agent.c): Manages conversation history (up to 32 messages with auto-compaction), implements a ReAct tool-calling loop (up to 10 turns), and preserves reasoning_content across multi-turn tool calls in thinking mode.
• Main Layer (mimo_main.c): CLI entry point supporting both interactive and single-shot modes.

Recommended Sampling Parameters

Per Xiaomi's documentation:

• top_p=0.95 (hardcoded)
• temperature=0.8: math, writing, web development
• temperature=0.3: agentic tasks (tool use)

Dependencies

• NETUTILS_CJSON: JSON parsing
• NETUTILS_WEBCLIENT: HTTP client
• CRYPTO_MBEDTLS: TLS encryption