M5Stack LLM630 Compute Kit (23.6g AXERA AX630C 12.8-TOPS INT4 Offline Edge-LLM & ESP32-C6 DevKit)
23.6-gram (90.3×31.6×12.0mm) ~3W fanless offline LLM and vision compute kit powered by the AXERA AX630C SoC (3.2 TOPS INT8 / 12.8 TOPS INT4 Transformer NPU), 4GB LPDDR4 (2GB dedicated to NPU), 32GB eMMC 5.1, onboard ESP32-C6 Wi-Fi 6 coprocessor, Gigabit Ethernet, BMI270 IMU, duplex audio, and M5 StackFlow plug-and-play AI.
Editorial Rating Methodology (★ 4.7 / 5.0): This score is an internal editorial evaluation compiled by the RECIWO Labs editorial team—not a user poll or crowd-voted rating. It assesses published silicon compute efficiency (25%), optical/sensor/actuator engineering (25%), battery & thermal envelope (20%), SDK/software openness (15%), and price-to-specification value (15%) based on official manufacturer datasheets and public technical documentation. Learn more about our evaluation rubric →
Key Engineering Highlights
- Runs complete offline voice-assistant pipelines (Wake-Word -> Whisper ASR -> Qwen2.5 / Llama-3.2-1B -> MeloTTS) inside ~3W of passive power using AXERA's 12.8-TOPS INT4 Transformer NPU
- Eliminates Linux ML cross-compilation headaches via M5Stack's open-source StackFlow framework—exposing plug-and-play JSON APIs over Grove UART, Gigabit Ethernet, or ESP32-C6 Wi-Fi 6
- Packs 4GB LPDDR4, 32GB pre-flashed eMMC 5.1, a BMI270 6-axis IMU, and full-duplex NS4150B speaker/mic interfaces into a 23.6-gram (90.3×31.6mm) module
Full Hardware Specifications
| Processor / Edge SoC | AXERA AX630C Edge AI SoC (Dual-Core ARM Cortex-A53 @ 1.2GHz) + Onboard ESP32-C6 Wi-Fi 6 Companion MCU |
|---|---|
| AI Model Stack | 3.2 TOPS @ INT8 / 12.8 TOPS @ INT4 Mixed-Precision NPU Running StackFlow (Qwen2.5, Llama-3.2-1B, Whisper, MeloTTS, YOLO11) |
| Display & Optics | MIPI DSI Display FPC Interface + Status LEDs + Headless JSON-RPC UART/TCP Control for M5Stack/Arduino Hosts |
| Camera & Vision | MIPI CSI Camera FPC Connector Supporting Low-Latency ISP + Hardware YOLO11 / InternVL2.5 Vision Pipelines |
| Measured Weight | 23.6g Bare Compute Kit (90.3 × 31.6 × 12.0mm Thumb-Stick Footprint; ~2.5W–3.5W Passive Fanless Operation) |
| Battery & Runtime | 5V USB Type-C Power Input (Dual USB-C OTG & Debug Ports + Ultra-Low ~3W Envelope for Portable LiPo Packs) |
| Connectivity & I/O | JL2101 Gigabit Ethernet + ESP32-C6 2.4GHz Wi-Fi 6 (w/ SMA Antenna Port) + 2× HY2.0-4P Grove Ports + Dual USB-C |
| Sensors & Acoustics | 4GB LPDDR4 (2GB System + 2GB Dedicated NPU) + 32GB eMMC 5.1 + MicroSD + BMI270 6-Axis IMU + NS4150B Audio Amp/Mic |
⚖️ Side-by-Side Comparison: M5Stack LLM630 Compute Kit (23.6g AXERA AX630C 12.8-TOPS INT4 Offline Edge-LLM & ESP32-C6 DevKit) vs. Edge Compute & DevKits Peers
Direct specification comparison against leading competing devices in the Edge Compute & DevKits category.
| Specification | M5Stack LLM630 Compute Kit (23.6g AXERA AX630C 12.8-TOPS INT4 Offline Edge-LLM & ESP32-C6 DevKit) (Current) | NVIDIA Jetson Orin NX Super Edge AI DevKit → | Xiaozhi AI Voice Cube Pro (ESP32-S3 Open DevKit) → | Apple Mac mini (M5 Pro 64GB) Private AI Node → |
|---|---|---|---|---|
| Brand | M5Stack | NVIDIA | Open Hardware Community | Apple |
| Retail Price | $99 | $499 | $12 | $1,799 |
| Editorial Rating | ||||
| Edge SoC | AXERA AX630C Edge AI SoC (Dual-Core ARM Cortex-A53 @ 1.2GHz) + Onboard ESP32-C6 Wi-Fi 6 Companion MCU | NVIDIA Ampere GPU (1024 CUDA + 32 Tensor Cores) | Espressif ESP32-S3R8 (Dual-Core LX7 @ 240MHz + 8MB PSRAM) | Apple M5 Pro (14c CPU + 20c GPU + 16c Neural Engine) |
| AI Model Stack | 3.2 TOPS @ INT8 / 12.8 TOPS @ INT4 Mixed-Precision NPU Running StackFlow (Qwen2.5, Llama-3.2-1B, Whisper, MeloTTS, YOLO11) | Runs Llama-3.2 8B / Qwen2.5-VL 7B / YOLO11 Locally | Edge ESP-SR Wake + Realtime Cloud Voice LLMs | MLX Accelerated Qwen3-32B / Llama-3.3-70B |
| Display / Optics | MIPI DSI Display FPC Interface + Status LEDs + Headless JSON-RPC UART/TCP Control for M5Stack/Arduino Hosts | HDMI 2.1 / DP 1.4 (Up to 8K) | 1.28-inch Circular IPS LCD (240×240) / 0.96" OLED | Supports up to three 6K displays |
| Camera | MIPI CSI Camera FPC Connector Supporting Low-Latency ISP + Hardware YOLO11 / InternVL2.5 Vision Pipelines | Dual MIPI CSI-2 Camera Lanes + 4× USB 3.2 | Optional OV2640 / GC2145 Camera Module | External USB-C / Studio Display Camera |
| Weight | 23.6g Bare Compute Kit (90.3 × 31.6 × 12.0mm Thumb-Stick Footprint; ~2.5W–3.5W Passive Fanless Operation) | 450g (including active heatsink & carrier board) | 62g (with 3D-printed enclosure & speaker cavity) | 730g |
| Battery | 5V USB Type-C Power Input (Dual USB-C OTG & Debug Ports + Ultra-Low ~3W Envelope for Portable LiPo Packs) | 19V DC Input (15W–40W configurable TDP) | 800mAh LiPo (5h continuous dialogue) | Internal 155W PSU (4W idle draw) |
📚 Related Engineering Teardowns & Reviews
Browse All 158 Articles →Orange Pi 5 Ultra ($119 89×56mm RK3588 LPDDR5, 4K60 HDMI-In & Wi-Fi 6E) vs M5Stack LLM630 ($99 23.6g AX630C 12.8-TOPS INT4 Offline LLM Kit) vs Radxa ROCK 5B+ Comparison: Compact SBC vs 3W NPU
A technical memory-bandwidth, video-capture, and sub-3W Transformer-NPU comparison of the Orange Pi 5 Ultra ($119, 60.5g 89×56mm RK3588 SBC packing a 6-TOPS NPU, 16GB LPDDR5, PCIe 3.0 x4 NVMe, 4K60 HDMI-In + 8K60 HDMI-Out, Wi-Fi 6E, and 2.5GbE), M5Stack LLM630 Compute Kit ($99.90, 23.6g AXERA AX630C devkit delivering 3.2 INT8 / 12.8 INT4 TOPS with 4GB LPDDR4, 32GB eMMC, ESP32-C6 Wi-Fi 6, and StackFlow offline Qwen2.5/Whisper/YOLO11), and Radxa ROCK 5B+ ($149).
Unitree B2 ($75k 60kg 360N·m, >40kg-Load, IP67 & 2,250Wh Industrial Quadruped) vs Hello Robot Stretch 3 ($24.95k 24.5kg Prismatic-Mast Dual-RealSense Open VLA Manipulator) vs Lynx M20 Comparison: Heavy Patrol vs Home VLA
A technical 360 N·m joint actuator, prismatic Cartesian mast vs serial arm kinematics, and IP67 field endurance comparison of the Unitree B2 ($75,000 tier, 60kg IP67 industrial quadruped delivering 360 N·m peak joint torque, >6.0 m/s speed, >=120kg standing / >40kg continuous walking load, and >5h / >20km range on a 2,250Wh battery), Hello Robot Stretch 3 ($24,950, 24.5kg 33×34cm mobile manipulator with 1.1m lift mast, 52cm telescoping arm, Dex Wrist 3, head D435i + wrist D405 RGB-D, Intel NUC12, and 100% open-source ROS 2 / LeRobot VLA stack), and DEEP Robotics Lynx M20 ($18,900).
Qualcomm RB3 Gen 2 Vision Kit ($399 QCS6490 12-TOPS Hexagon NPU & IMX577/OV9282 Dual-Cam) vs Khadas Edge2 Pro ($239 25g 5.7mm RK3588S2 6-TOPS 16GB SBC) vs RUBIK Pi 3 Comparison: Calibrated VIO vs Ultra-Thin SBC
A technical 6nm Hexagon 770 Tensor NPU vs 8nm RK3588S2 NPU, factory-calibrated VIO/SLAM camera mezzanine vs 5.7mm FPC density, and Linux toolchain comparison of the Qualcomm Dragonwing RB3 Gen 2 Vision Development Kit ($399, ~200g QCS6490 8-core Kryo 670 + 12-TOPS NPU + Triple ISP kit bundling a 12MP IMX577 + global-shutter OV9282 Vision Mezzanine, Wi-Fi 6E, and Qualcomm Linux / ROS 2 / AI Hub), Khadas Edge2 Pro ($239, 25g 82×57.5×5.7mm credit-card SBC with RK3588S2 6-TOPS NPU, 16GB LPDDR4X, 64GB eMMC, 3× CSI, 4× displays, and OOWOW), and Thundercomm RUBIK Pi 3 ($119).
XREAL One ($499 84g X1 Spatial Chip, 3ms 3DoF & 50°-FOV Bose Micro-OLED) vs RayNeo Air 2s ($349 78g 600-Nit 120Hz & Push-Push Quad-Speaker) vs XREAL One Pro Comparison: Native 3DoF vs Acoustic XR
A technical motion-to-photon (M2P) spatial silicon, birdbath vs flat-prism Micro-OLED optics, and phase-canceling open-ear acoustics comparison of the XREAL One ($499, 84g tethered AR glasses integrating the custom X1 spatial co-processor for 3ms native 3DoF anchor pinning, dual 0.68\" Sony 1080p 120Hz Micro-OLED panels at 50° FOV / 600 nits, 3-stage electrochromic dimming, Sound by Bose, and modular 12MP XREAL Eye AI camera slot), RayNeo Air 2s ($349, 78g 0.55\" Sony Micro-OLED 46° FOV / 600-nit glasses with Push-Push symmetric quad speakers), and XREAL One Pro ($599).
