Small AI models — those with a few billion parameters or fewer — are enabling life-saving applications in regions with no reliable internet, data centers, or stable electricity. Real-world examples include a handheld pill-authentication spectrometer running on an Android phone in Africa, drone-based crop disease detection in India, malaria mosquito detection, and ECG monitoring on Arduino devices in Brazil. These models are created via pruning, distillation, or quantization of larger models, and increasingly run on commodity hardware like a $50 Arduino with a Qualcomm chipset drawing just 3 watts. Open-weight models like Google DeepMind's Gemma 4 and Alibaba's Qwen 3.5 are accelerating adoption. The World Bank now actively funds small AI development globally. Advocates argue that millions of small, specialized edge models — not one giant centralized model — represent the sustainable future of AI for most of humanity, though infrastructure challenges remain.
Nguồn: https://spectrum.ieee.org/small-language-models-ai-pharmaceuticals. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.
Mekotronics has launched an AI Box based on NVIDIA Jetson Orin Nano (4GB/8GB) and Jetson Orin NX (8GB/16GB) modules, targeting humanoid robots, smart transportation (V2X), Smart Cities, Smart Agriculture, and medical imaging. The compact embedded computer (136×79mm PCB) offers standard industrial interfaces including HDMI, six USB Type-A ports, dual Gigabit Ethernet (one with PoE), a 40-pin GPIO header, CAN Bus, UART, two camera connectors, and dual M.2 NVMe slots. It supports up to 157 TOPS with the Orin NX 16GB module and runs JetPack 7.x on Ubuntu 24.04. The design appears cost-optimized compared to rugged competitors like IBASE EC3100 and Forlinx FCU3011, though pricing has not yet been announced.
Running the largest model that fits on your GPU isn't always the best strategy. Small 2B models like Gemma 4 E2B and Qwen 3.5 2B, purpose-built for edge hardware, handle the majority of everyday tasks — explaining concepts, summarizing, image analysis, structured text tasks — without maxing out VRAM or squeezing other processes. These models aren't stripped-down versions of larger ones; they're architected from the ground up for efficiency on consumer hardware, with features like 128K context windows, multimodal input, and tool calling.
The Titan Mini is a compact, lower-cost development board based on the Renesas RA8P1 (R7KA8P1) microcontroller, featuring a 1 GHz Arm Cortex-M85 core, a Cortex-M33 at 250 MHz, and an Arm Ethos-U55 NPU delivering 256 GOPS. Compared to the original Titan board, it reduces memory to 32MB SDRAM and 8MB eMMC, replaces the RJ45 Ethernet port with an FPC connector, and adds a built-in microphone and speaker connector. The board measures 65×55mm and includes a 40-pin Raspberry Pi HAT-compatible GPIO header, USB-C, CAN Bus, JTAG, and optional camera support. RT-Thread provides an SDK and BSP on GitHub with demos including LED blink, display/camera, Ethernet, NPU-accelerated face detection, and a WAV audio player. It sells for $44.23 on AliExpress (Ethernet cable extra), but is not shipped to the EU due to unspecified regulatory/certification issues.
Vision AI agents are increasingly used to extract operational intelligence from video data in factories, cities, and warehouses. Three common blockers — data gaps causing accuracy plateaus, lack of fine-tuning expertise, and complex agent assembly — can be addressed using NVIDIA Metropolis blueprints, NVIDIA Omniverse for OpenUSD-based synthetic data generation, and NVIDIA TAO for model fine-tuning. Three real-world workflows illustrate the approach: Roboflow and Corning used synthetic defect images to achieve 95% average precision from just eight real training samples; Linker Vision cut smart city development effort by 85% and incident response times by 80% using the VSS blueprint; and DeepHow's SOP Verification agent at Foxconn improved first-pass yield by 3% and achieved 99% task-level accuracy on assembly line operations.
Makerfabs has released the MaTouch ESP32-P4 TFTTouch 10.1, a development board combining the ESP32-P4 and ESP32-C6 chips with a 10.1-inch 1280×800 IPS touchscreen display. Key differentiators from similar boards include a SIMCom SIM7670G 4G LTE Cat 1 modem with SIM slot, RJ45 Ethernet, and a 2MP MIPI camera. The board also includes WiFi 6, Bluetooth 5, Zigbee/Thread/Matter via the ESP32-C6, audio I/O, microSD, five USB-C ports, and GPIO expansion headers. Software support covers ESP-IDF, Arduino IDE, PlatformIO, and LVGL. It is priced at $79.80 and ships with a 32GB microSD card in a basic enclosure.
Onsemi is acquiring Synaptics in an all-stock deal valued at approximately $7bn (including debt), betting that AI's next growth wave will happen at the edge — in cars, factories, and robots — rather than in cloud data centers. Onsemi brings power and sensing chips; Synaptics contributes connected compute via its Astra platform with AI processors, NPUs, and wireless connectivity. The combined company aims to cover every layer of the Edge AI stack. The deal carries a ~19% premium for Synaptics shareholders, who will own about 12% of the merged entity. Onsemi expects $200m in annual savings and a $30bn expansion of its addressable market by 2030. However, investors reacted cautiously — onsemi shares fell 8.2% after hours — citing the company's failed $6.9bn Allegro MicroSystems bid a year prior, planned job cuts, and a mid-2027 closing timeline subject to regulatory review.
MYiR has launched the MAC-B5760, a fanless Edge AI industrial box PC built around the Rockchip RK3576 SoC featuring a 6 TOPS NPU. It ships with 8GB LPDDR5 and 64GB eMMC by default, offering dual Gigabit Ethernet, four USB 3.0 ports, HDMI 2.1, Mini DisplayPort, optional WiFi 6/BT 5.4, and expansion via M.2 and mini PCIe slots. An optional Rockchip RK1828 M.2 AI accelerator module (20 TOPS) supports LLM/VLM inference for models up to 7 billion parameters. Software support includes a Linux 6.1 BSP with Debian, Yocto, and Yocto Preempt RT images. The standard model starts at $299, while the variant with the RK1828 AI module is $599.
Graperain has announced the GR1126MB development board, built around a 42x42mm GR1126B stamp-hole system-on-module powered by the Rockchip RV1126B SoC. The SoC features a quad-core Cortex-A53 CPU, a 3 TOPS NPU supporting TensorFlow, ONNX, PyTorch, and Caffe, plus a 12MP ISP and AI-ISP. The module offers up to 4GB LPDDR4x RAM and 256GB eMMC, with extensive I/O including dual MIPI CSI, Gigabit Ethernet, USB 3.0, CAN Bus, and more. The development board adds WiFi/BT, optional 4G LTE, a 40-pin GPIO header, and dual camera connectors. Software support includes BuildRoot and Debian 12, along with an AI framework for object detection, face recognition, and edge analytics. No pricing has been disclosed.