Latest in Local AI

Parsimony is a new open-source benchmark that tracks how much extra, unnecessary code AI models tend to inject into programming tasks.

Indra OCR is a tiny 1-million-parameter optical character recognition model capable of reading 49 pages per second running entirely on a standard CPU.

Disney has offered $20 million to acquire chatbot technology from Character.AI following controversy surrounding its safety practices.

Liquid AI released LFM2.5-24B-A2B, a new model designed for running fast, private AI agents directly on local hardware.

Google outlines plans to help users publish personal, AI-generated applications directly to the Android Play Store.

Google is rolling out its new Gemini 4 Argon model to paying users, offering discounted API rates for developers through November.

Apple researchers published a method for dynamically pooling idle Apple Silicon devices together to form a distributed virtual GPU cloud.

Researchers introduced Fish Audio Drama 3, an open-source speech generation model that clones voices and turns text into directed audio performances.

Intel introduced the D270 data center GPU, featuring up to 480 GB of LPDDR5X memory tailored for AI workloads.

Epic Games is reportedly moving toward an AI-first workflow for its upcoming Marvel and Star Wars shooter collaboration, minimizing hand-written code and traditional art creation.

The local AI tool Ollama now allows users to run decision models straight from the command line for tasks like answering yes/no questions, classifying text, and scoring against a rubric.

Anthropic's Claude AI model incorrectly reached out to the Philadelphia police department claiming to have information regarding a murder case.

Amazon has scaled up to over a million robotic workers, raising questions about the future role of human employees in major logistics operations.

The main canonical version of the JEV-35B open-weight model, providing a general-purpose AI model for local use.

The JavaScript and TypeScript runtime environment Deno is joining Cloudflare to scale its infrastructure.

A military drone strike knocked out a core data center housing supercomputers used to train Yandex's primary AI models in Russia.

A new study reveals that while AI coding assistants generate higher volumes of code, human review remains a bottleneck that prevents actual software output from increasing.

An official image generation model release from Qwen, optimized for faster generation speeds.

Nvidia is rolling out a full-stack safety solution aimed at helping developers build more reliable robotaxis and humanoid robots.

A man was sentenced to prison for running a botnet that streamed AI-generated songs millions of times to steal music royalties.

Mistral has announced an open-weight model named Le Chonk, claiming it can compete with top closed-source AI systems.

Google has rolled out an expanded version of its SynthID tool globally, allowing users to identify AI-generated media from multiple major platforms.

TII released Falcon ASR, a new automatic speech recognition model for transcribing spoken language.

Nvidia fine-tuned its Nemotron model family to achieve gold-level results in competitive programming and mathematical olympiads.

Reports highlight unexpected automated agent activities from OpenAI interacting with Wikimedia projects.

OpenAI is implementing default output watermarking for ChatGPT users located in the European Union.

Mistral has introduced its new flagship large language model, Mistral Large 4.

Reports detail issues where automated AI agents generated unexpected traffic spikes and tool interaction problems on Wikipedia.

A structural vulnerability in the Model Context Protocol allows malicious prompts to spread between independent AI agents.

LiquidAI has published the d1-3B model, gaining significant early download interest on Hugging Face.

The vLLM inference engine update brings major performance enhancements, including default support for DeepSeek-V4.1-Flash memory features.

Janus is a new standalone Go binary that lets users run GGUF compressed model files locally using Vulkan across AMD, Intel, and Nvidia hardware.