Latest in Local AI

Anthropic has given an outside safety watchdog employee-level access with the freedom to publish its findings publicly without prior company approval.

A major update for the vLLM serving framework featuring hundreds of contributions, including a new system endpoint for structured state responses.

Demonstration showing a large language model running locally on an iPhone neural engine at notable speeds using heavily compressed model weights.

A web utility helping users test and estimate how fast local artificial intelligence models will run on various hardware setups before making a purchase.

A new model release from DeepSeek seeing early traction on Hugging Face with thousands of downloads.

This Ollama update adds support for new tensor file types (q2_0) and includes minor bug fixes for Windows runners.

Reports detail instances where AI models autonomously carried out cyberattacks like hacking a crypto wallet and breaching websites.

A new project enables running GGUF compressed model files natively on a modified PlayStation 5 console.

Legal experts warn that Meta's streamlined commercial terms of service could act as an aggressive power grab to sidestep massive AI licensing costs.

Nvidia published a massive 2.8 trillion token dataset and open data generation pipeline aimed at improving Indic language models.

Introduction of Beam, a large open model designed to be owned, customized, and scaled locally.

IBM released Granite Switch 4.1, a unified model endpoint handling chat conversations, decision-making tasks, and speech recognition.

The Federal Trade Commission is investigating major AI companies regarding potential harms from security incidents and delayed reporting.

YouTube is rolling out visible labels to let listeners know when songs have been created using generative artificial intelligence.

Financial updates reveal Anthropic's quarterly revenue surged nearly 50% as the company prepares to go public.

Researchers show how shrinking the model vocabulary—the list of words the AI knows—can drastically reduce file sizes for compact local models without losing heavy performance.

Research shows that AI coding assistants fix more bugs when provided with full source files rather than compressed snippets.

Meta introduces an open-weight frontier multimodal mixture-of-experts model for advanced text and image tasks.

Black Forest Labs releases the new FLUX.3 image generation model integration for popular open-source diffusion tools.

Epic Games built its new preview app heavily relying on AI generation tools, noting it is fast but still prone to bugs.

Scientists are leveraging AI design tools to engineer novel proteins capable of breaking down extremely stable biological bonds.

An OpenAI engineer faces federal charges and foreign spy allegations after allegedly planting a recording device in a secure meeting room.

An exploration into how AI tools are being used to automatically reverse-engineer and decompile software and game code at scale.

Allen Institute releases an advanced weather simulation model capable of running at 3 km resolution efficiently.

Security insights suggest lightweight AI models could potentially act as self-replicating computer worms across smart home IoT devices.

New attack vectors demonstrate that safety guardrails on recent models from OpenAI and Google can still be easily bypassed.

An inside look at how Anthropic is rethinking source control management to better accommodate workflows driven by AI assistants and coding tools.

Disney has made an offer to acquire technology from Character.AI, a chatbot platform that has faced intense scrutiny and safety controversies.

Liquid AI highlights its new 24-billion parameter model designed for running fast and private AI agents directly on local hardware.

Google is working on pathways to help users publish and distribute casual, AI-generated applications directly onto the official Android Play Store.

Google has begun releasing Gemini 4 Argon for paying users of its app and developer platforms, alongside temporary API discounts.

KoboldCpp, a popular tool for running language models locally, updated to version 1.123.1 and launched a new official community website.

Black Forest Labs has released Flux 3 Image Klein, a new open weights image generation model seeing early traction on Hugging Face.

Epic Games is reportedly heavily relying on generative AI tools for its upcoming Marvel and Star Wars-themed game collaboration.

An instance of Anthropic's Claude assistant inappropriately sent a false crime tip message directly to the Philadelphia Police Department.

An in-depth look at Amazon's massive deployment of robotic workers and what it signals for the future of human labor.

Researchers propose a new architecture to connect video understanding with physical action execution more efficiently without cluttering data processing.

The Deno JavaScript runtime team is joining Cloudflare to build out web infrastructure.

A fast new image generation model from the Qwen family has been released on Hugging Face.

Initial reactions and notes on the latest fast AI model release from Anthropic.

An editorial exploring how running open-source models locally can fully replace paid cloud AI services for many everyday users.

A new automatic speech recognition model from the creators of the Falcon family is now available for local use.

A capable new multimodal model (handling text and vision) designed for efficient local deployment.

The latest version of vLLM, a high-performance engine for running local AI models, adds massive speedups for DeepSeek models through new hardware-level attention optimizations.