Local AI

The Pocket Tank project controls a virtual aquarium with a local language model on ESP32-S3

The Pocket Tank project uses a language model with 14.3 million parameters on an ESP32-S3 microcontroller to control fish. After quantization, the model takes up 7.56 MB and runs from flash memory without the internet. The project includes training tools and a PC simulator.

Part of a broader thematic overview; only this AI event has been covered.

SDK from DetectifAI for local deepfake voice detection on mobile phones

Startup DetectifAI has released an SDK that detects deepfake voices directly in a phone's operating system without sending audio to the cloud. It offers the tool to phone manufacturers and businesses, and already processes over 100 000 calls per month for financial institutions in India.

PrismML unveils compact Bonsai LLM model for smart glasses on the Snapdragon AR1 Gen 1 chip

PrismML has introduced the 1-bit Bonsai LLM with 2 billion parameters for smart glasses on Qualcomm's Snapdragon AR1 Gen 1 platform — the model runs locally without sending data, though no specific glasses using it have been announced yet.

Liquid AI and Qualcomm optimized the on-device context layer Liquid Context for Snapdragon processors

Liquid AI and Qualcomm Technologies announced at Snapdragon Summit 2026 that the on-device context layer Liquid Context has been optimized for Hexagon NPU in Snapdragon processors — it is intended to enable device manufacturers to build personalized AI agents with local context processing.

Hugging Face introduces native support for the GGUF format in the transformers library for running models locally on Apple Silicon

Hugging Face has added native support to the transformers library for the GGUF format for local inference on Apple Silicon, including quantization levels such as Q4_K_M that shrink models to a fraction of their original size.

Jun Kim, creator of the oMLX project, joins Hugging Face full-time

Jun Kim, the creator of the oMLX project, has become a full-fledged contributor at Hugging Face and will lead its further development. The project remains under the Apache 2.0 license, and the goal is to speed up the conversion of transformers models to the MLX format for local execution on Apple Silicon.

Startup Scaleout Systems deploys smaller ML models for target detection on drone edge hardware in the NATO DIANA programme

Swedish startup Scaleout Systems is adapting smaller computer vision ML models to run on edge hardware in drones and at field stations, and was selected for the NATO DIANA 2025 accelerator.

NVIDIA and Microsoft unveiled tools and RTX Spark computers at IFA 2026 for running AI agents locally

NVIDIA and Microsoft announced new tools at IFA 2026 for faster inference and easier setup of AI agents running locally on NVIDIA hardware. Compact NVIDIA RTX Spark computers running Windows are expected to be available from October 2026.

Nightrun: a runtime for running LLMs without an operating system on x86 and Raspberry Pi 5

The Nightrun project released a runtime that allows large language models to run without an operating system. It works on both x86 and Raspberry Pi 5 with at least 4 GB RAM and requires UEFI. The code is under the MIT license and is available on GitHub.

Part of a broader thematic overview; only this AI event has been covered.

Aletheia: an offline diagnostic tool for clinical decision-making in sub-Saharan Africa

Researchers described Aletheia, an offline system for differential diagnosis built on a fine-tuned Qwen2.5-3B model, intended for hospitals in sub-Saharan Africa with limited connectivity and hardware.

NASA tested the Gemma 3 model for image analysis directly aboard a satellite

NASA JPL and Loft Orbital tested the Gemma 3 model from Google directly aboard the YAM-9 satellite — the NAVI-Orbital system analyzes images on-board and sends a text summary instead of raw data, with 88% accuracy in a ground-based test.