join.dev
AI news · 10 stories
Google DeepMind and GovTech develop on-device scam detection · Liquid AI releases LFM2.5 for local reasoning agents · NEURA introduces Neuron AI for multisensory robot control
Google DeepMind and Singapore’s GovTech developed a Gemma-based model trained on local scam conversations to detect manipulation on smartphones. Processing stays on the device, offering a way to flag scams while preserving conversation privacy.
Read at Google DeepMind
Liquid AI released open-weight LFM2.5-24B-A2B and its base model for local reasoning and tool-use workflows. The mixture-of-experts model activates roughly 2B parameters per token, aiming to make persistent agents practical on personal hardware.
Read at Hugging Face
NEURA introduced Neuron AI, combining vision, touch, audio, robot state, language and memory in a shared robotics model. It supports five NEURA robot embodiments, with task fine-tuning and fleet deployment through Neuraverse.
Read at NEURA Robotics
Danaher announced a collaboration combining its scientific instruments and laboratory workflows with NVIDIA’s accelerated computing and AI. The partners aim to connect model predictions with experimental evidence in a continuous learning cycle.
Read at Danaher
Ai2 introduced HiRO-ACEv2, which combines a coarse climate emulator with downscaling for rain, wind and pressure. The team reports global 3 km climate simulations at about one three-hundredth the cost of X-SHiELD, expanding options for studying extreme weather.
Read at Hugging Face
Moondream announced its acquisition by Prime Intellect, joining its open vision models and Photon inference technology with Prime’s training stack. Moondream says its models will remain open, while its cloud inference and Lens services will shut down November 13 with migration support.
Read at Moondream
Deezer announced a toggle due before year-end that will let listeners exclude fully AI-generated music from search results. The feature extends its existing detection, labeling and recommendation exclusions, giving listeners more control over discovery.
Read at Deezer
NVIDIA outlined city projects using its models and computing stack for infrastructure monitoring, public services and digital twins. The initiatives include operational deployments and exploratory pilots, showing how agents are being adapted to local government workflows.
Read at NVIDIA
DMIS Lab released UltraBERT English encoders in 150M and 380M sizes with native 32K context, plus a ready-to-use embedding model. Apache 2.0 licensing offers developers compact building blocks for classification and document retrieval.
Read at Hugging Face
Square unveiled an updated self-service kiosk with conversational voice ordering grounded in a restaurant’s menu and sales data. Customers can ask questions and modify orders through voice or touch, extending AI assistance into restaurant checkout.
Read at Digital Transactions