This week brings a surprisingly capable AI experiment for the ESP32-S3, an industrial Raspberry Pi Compute Module gateway, and a compact Jetson AGX Orin system built for robotics and multi-camera AI.

There's a huge gap between an $8 microcontroller and a $2,000+ edge AI computer, but both are being pushed in the same direction: doing more locally.

An ESP32-S3 is now running a 28.9-million-parameter language model entirely on-device, RAKwireless has brought the Raspberry Pi Compute Module 5 into a rugged outdoor IoT gateway, and Seeed Studio has packed a Jetson AGX Orin 32GB into a compact system aimed at robotics and computer vision.

Running AI models on microcontrollers isn't new. TinyML projects have handled wake-word detection, image classification, and sensor analysis for years.

Running a 28.9-million-parameter language model on an ESP32-S3 is a little different.

Developer slvDev's open-source esp32-ai project generates text entirely on an ESP32-S3 with just 512KB of SRAM, 8MB of PSRAM, and 16MB of flash. Nothing is sent to a cloud server, and the project reports an end-to-end generation speed of 9.88 tokens per second.

The model itself is 14.9MB after 4-bit quantization, which makes fitting it into the ESP32's limited memory the real challenge.

The solution is to avoid treating every parameter equally.

Of the model's 28.9 million parameters, about 25 million are stored in a flash-based lookup table. Frequently accessed data, including activations and normalization weights, stays in the ESP32-S3's much faster SRAM, while the model's dense core and output head sit in PSRAM.

The large embedding table remains in flash, and only the small portion needed for each generated token is retrieved. According to the developer, that works out to around 450 bytes per token.

The technique is based on Google's Per-Layer Embeddings, also used in Gemma 3n, but adapted here around the ESP32-S3's much tighter memory constraints.

There is an important limitation: this isn't a general-purpose chatbot squeezed onto an ESP32.

The demonstration model was trained on TinyStories and mainly produces short, simple stories. It can't reliably answer general questions, write code, or behave like an instruction-following assistant.

But that's not really what makes the project interesting.

The bigger achievement is showing that a transformer model of this size can run locally on hardware this inexpensive. Similar techniques could eventually be useful for offline command interpretation, predictive text, simple device assistants, or other narrowly trained models that need to work without an internet connection.

For an $8 microcontroller, 28.9 million parameters at nearly 10 tokens per second is an impressive place to start.

RAKwireless' RAK7393 WisGate Connect Outdoor shows a very different direction for Raspberry Pi hardware.

Rather than building another desktop-style SBC around the Raspberry Pi Compute Module, the RAK7393 turns either a Compute Module 4 or Compute Module 5 into the center of a rugged industrial IoT gateway.

The system includes three Mini PCIe slots that can be populated with different radio modules. RAK supports configurations such as a sub-GHz LoRaWAN concentrator alongside a 2.4GHz LoRaWAN radio, as well as combinations involving LoRaWAN, mioty, and cellular backhaul.

That flexibility allows the RAK7393 to act as a multi-protocol edge gateway built around the radios a particular deployment actually needs.

The rest of the hardware is equally industrial-focused. It includes Gigabit Ethernet with 802.3at PoE, 9-36V DC input, RS-485, GNSS, a real-time clock, and an ATECC608 cryptographic chip.

Three onboard supercapacitors provide enough backup power for what RAK calls "dying-gasp" functionality, allowing the system to detect a power failure and perform limited shutdown or reporting tasks before losing power completely.

Everything is housed inside an IP67/NEMA-6 enclosure rated for temperatures from -20°C to 85°C and designed for wall or pole mounting.

CM5 support is what makes the RAK7393 particularly interesting.

Moving from the CM4's Cortex-A72-based BCM2711 to the CM5's 2.4GHz quad-core Cortex-A76 BCM2712 provides a substantial jump in CPU performance. RAK supports CM5 configurations with up to 16GB of RAM and 64GB of onboard eMMC storage.

RAK also provides its own RAKPiOS, based on Raspberry Pi OS, with the necessary drivers, security additions, helper scripts, and Docker already included.

That means the RAK7393 can do considerably more than forward LoRaWAN packets. It can run containerized applications and perform processing directly at the edge.

It's a good example of how far Raspberry Pi's Compute Module lineup has moved beyond the usual idea of a Raspberry Pi.

Instead of sitting on a desk connected to a monitor, a CM5 can now be mounted outside on a pole, powered over Ethernet, and spend years operating as part of an industrial sensor network.

If the ESP32 project shows how far AI can be scaled down, Seeed Studio's reComputer Mini J5011 with GMSL shows what happens when much more computing power is packed into a relatively small edge system.

At its center is NVIDIA's Jetson AGX Orin 32GB, giving the J5011 up to 200 TOPS of AI performance.

This isn't meant to compete with an ordinary desktop mini PC.

Seeed is targeting robotics, autonomous machines, humanoid development, and multi-camera computer vision, where the Jetson's GPU and AI accelerators can process large amounts of sensor data locally without constantly sending it back to a server.

The system measures roughly 119 × 119 × 100mm and packs in a 10GbE port alongside Gigabit Ethernet, two 10Gbps USB 3.2 Type-A ports, HDMI 2.1, a 128GB NVMe SSD, CAN, RS-485, I2S, and digital I/O.

One of its most interesting features is GMSL2 camera support.

Two FAKRA connectors allow the system to support up to eight GMSL2 cameras simultaneously, making it particularly useful for robots and autonomous systems that need multiple viewpoints for navigation, object recognition, mapping, or depth perception.

The system also includes two CAN-FD interfaces, another feature that makes far more sense on a robotics platform than another collection of standard desktop ports.

On the software side, Seeed ships the system with JetPack 6.2 and also lists support for JetPack 7.2, alongside frameworks and tools including Isaac ROS, ROS 2, PyTorch, and Hugging Face.

The complete reComputer Mini J5011 with GMSL and Jetson AGX Orin 32GB is currently listed at $2,899 as a limited-time early-bird price.

That puts it far outside normal maker SBC territory, but this is also a very different class of hardware. Its real competition is closer to industrial robotics computers and NVIDIA's own high-end Jetson development platforms.

What makes the J5011 interesting isn't just the 200-TOPS figure. It's the combination of Jetson AGX Orin performance, high-speed networking, industrial interfaces, and support for large multi-camera setups in a system compact enough to integrate directly into a robot or autonomous machine.

These three projects could hardly be more different.

At one end, an ESP32-S3 with only a few megabytes of memory is generating text locally using an architecture built specifically around its hardware limitations.

In the middle, Raspberry Pi's Compute Module 5 is moving deeper into industrial deployments where Linux, Docker, and multiple wireless networks need to run reliably outdoors.

At the high end, Seeed's reComputer Mini J5011 brings 200 TOPS of Jetson performance, industrial I/O, and multi-camera support into a compact system designed to become part of the machine itself.

The interesting trend isn't simply that everything is getting an "AI" label.

It's that local computing is becoming practical across an increasingly wide range of hardware — from an $8 ESP32 generating text on-device to a multi-camera Jetson system capable of controlling and interpreting the world around a robot in real time.