M5Stack Module LLM Kit – Offline AI Inference and Voice Assistant Module with Ethernet and USB Debugging

M5Stack has introduced the Module LLM Kit, a smart modular solution designed for offline AI inference and data communication in various applications, such as smart home devices, voice assistants, and industrial control systems.

It integrates the Module LLM and Module 13.2 LLM Mate modules to provide efficient and intelligent interaction without relying on the cloud, ensuring privacy and security. Key features include offline inference with 3.2 TOPS at INT8 precision, built-in functions like wake word detection (KWS), speech recognition (ASR), text-to-speech (TTS), and large language model (LLM) support. It offers 4GB LPDDR4 memory, 32GB eMMC storage, and serial communication capabilities. The system supports multi-model processing, uses an advanced AiXin AX630C SoC processor that includes a dual Cortex A53 @ 1.2 GHz, and has a low power consumption of around 1.5W. A variety of interfaces for system integration, such as USB, Ethernet, and serial communication, are also featured by the kit. It is perfect for offline voice assistants, smart home control, interactive robots, and text-to-speech conversion.

Building on the earlier M5Stack LLM Module, the Module LLM Kit adds the Module 13.2 LLM Mate, offering extra connectivity like an RJ45 Ethernet port, core serial support for SBCs, and USB-to-serial debugging through the CH340N chip.

M5Stack Module LLM Kit Specifications:

• SoC: AiXin AX630C
  • CPU: dual Cortex A53 @ 1.2 GHz
  • NPU: Max 12.8 TOPS @INT4, 3.2 TOPS @ INT8
• Memory: 4GB LPDDR4 RAM (1GB for applications, 3GB for hardware acceleration)
• Storage:
  • 32GB eMMC 5.1 storage
  • MicroSD slot
• Audio:
  • MSM421A microphone
  • AW8737 audio driver
  • 8Ω 1W, 2014 cavity speaker
• Connectivity:
  • 1x RJ45 Ethernet port (100 Mbps with onboard transformer)
  • FPC-8P (connects directly to Module LLM)
• Supported Models (via apt):
  • LLMs: Qwen2.5-1.5B, Llama-3.2-1B, deepseek-r1-distill-qwen-1.5b, InternVL2_5-1B-MPO
  • TTS: melotts
  • STT: whisper-tiny, whisper-base
  • Visual: yolo11 and other SOTA models
  • ASR
• USB:
  • USB Type-C (power and log output)
  • USB-to-Serial (CH340N chip)
  • Supports USB log output and ADB debugging
• Interfaces:
  • Serial communication, default baud rate 115200@8N1 (adjustable)
  • 3x RGB LED@2020, driven by LP5562 (status indicator)
  • HT3.96 9P solder pad (DIY expansion)
  • M5BUS connector (power and stacking)
  • UART for Tx and Rx
• Misc:
  • 3x RGB LEDs driven by LP5562
  • 1x Button (firmware download mode)
  • CH340N Conversion Chip
• Power Consumption: ~1.5W
  • No Load: 5V @ 0.5W
  • Full Load: 5V @ 1.5W
• Dimensions: 54.0 x 54.0 x 19.7 mm
• Weight: 19.2g
• Operating Temperature: 0°C to 40°C

The official apt repository includes a wide selection of AI models like deepseek-r1-distill-qwen-1.5b, InternVL2_5-1B-MPO, and Llama-3.2-1B. It also features text-to-speech models such as whisper-tiny and melotts, along with visual models like yolo11 for handling image-based tasks.

In terms of software support, the Module LLM Kit works with UiFlow1, UiFlow2, and the Arduino IDE, making use of the StackFlow framework and libraries to simplify AI development. For detailed technical info, check out the M5Stack documentation page.

At the time of writing, the M5Stack Module LLM Kit is available on the M5Stack store at $49.90 and on AliExpress at $51.66.