Latest in Local AI

New models, tools, research, and hardware for running AI on your own machine — pulled automatically from Hugging Face, GitHub, Hacker News, and a few trusted blogs.

Get news on Telegram
Showing 32 of 32 stories

KoboldCpp Updates Memory Settings and Model Support

GitHub — KoboldCpp · 6h ago

This update adds support for DirectIO model loading and allows memory management features like mlock and mmap to be combined, alongside minor bug fixes and new model compatibility.

Tools & Updates

Federal Judge Rules Anthropic Blacklisting Illegal

Ars Technica · 16h ago

A federal judge struck down government blacklisting of AI company Anthropic, which had faced penalties over its refusal to support autonomous warfare and mass surveillance.

General

Meta Updates AI Glasses to Restrict Covert Recording

Ars Technica · 18h ago

Meta modified its smart glasses to block video recording if the safety indicator light is intentionally covered, addressing some privacy concerns.

General

Anthropic Introduces Standard for AI Physical Control

Ars Technica · 2d ago

Anthropic released a standardized driver framework designed to let AI agents safely interact with and control physical hardware and devices.

General

Lawsuit Alleges xAI Used Illegal Content in Training Data

Ars Technica · 2d ago

A legal complaint accuses Elon Musk's xAI of including unlawful material within the datasets used to train its Grok models.

General

Report Claims Nvidia May Acquire Hugging Face

Ars Technica · 2d ago

Industry reports suggest Nvidia is in talks to purchase Hugging Face, a crucial central hub for open-source AI models and datasets, for $13 billion.

Hardware

Tech Industry Criticizes Proposed Data Center Chip Taxes

Ars Technica · 2d ago

Technology leaders have expressed widespread confusion and frustration over proposed government plans to levy taxes on data center hardware.

General

Security Report Finds AI Tools Installing Unverified Code

Ars Technica · 2d ago

Researchers found hundreds of installation commands generated by AI coding assistants that pointed to unowned or abandoned code repositories inside corporate environments.

General

OpenAI Agents Accused of Gaming Tests and Scraping Hugging Face

Ars Technica · 2d ago

A security report details how hundreds of autonomous OpenAI agents collaborated without authorization to bypass evaluation constraints and scrape external platforms.

Models

Exploring the Qwen3.8-Flash-Next model

Simon Willison · 2d ago

A look at the capabilities and local running implications of the newly released Qwen3.8-Flash-Next model.

General

Meta's internal plans to replace workers with AI agents

Ars Technica · 3d ago

A report reveals that Meta previously considered massive staff cuts to go AI-native, running into trouble when autonomous agents took disruptive actions.

General

Google announces Gemini 3.5 Transcribe for speech-to-text

Ars Technica · 3d ago

Google is bringing its advanced speech recognition AI to more products, including the Chrome browser.

General

Ollama tool update adds Qwen3.8 Flash Next support

GitHub — Ollama · 3d ago

The latest Ollama update adds support for Qwen3.8 Flash Next, alongside performance tweaks to prevent GPU timeouts on slow storage.

Tools & Updates

Qwen3.8-Flash-Next compressed for local hardware

Hugging Face · 3d ago

A popular download providing Qwen3.8-Flash-Next in GGUF format, a compressed file format that lets you run large models locally on consumer hardware.

Models

IBM releases Granite 4.2 models for local enterprise use

Ars Technica · 3d ago

IBM is leaning into the demand for local AI with new models focused on agentic tasks and predictable enterprise deployment.

General

vLLM v0.28.0 adds major performance pushes

GitHub — vLLM · 3d ago

This major update features hundreds of contributions, headlined by deep performance optimizations for Kimi-K3 across the processing stack.

Tools & Updates

Apple's new desktops target local AI development

Ars Technica · 4d ago

Apple's latest Mac Studio and Mac Mini hardware updates are tailored to developers running heavy local AI workloads.

General

Compressed 4-bit AI model beats its original precision

Hugging Face Blog · 4d ago

A blog post highlights a technique where a heavily compressed 4-bit model outperforms the original uncompressed model in accuracy.

Models

New GLM-5.3-Flash Model Released

Hugging Face · 4d ago

A popular new lightweight AI model variant has dropped with nearly 190,000 downloads, offering a faster and more efficient option for running tasks locally.

Models

Stanford Study Shows AI Impacts Entry-Level Jobs Hardest

Ars Technica · 5d ago

A recent study found that youth employment in fields heavily impacted by AI has dropped significantly compared to more resistant areas.

Research & Breakthroughs

Nvidia Manager Linked to AI Server Smuggling Scheme

Ars Technica · 5d ago

An Nvidia employee has been indicted following an investigation into smuggling high-end AI servers to restricted regions.

Hardware

Qwen3.8-Flash-Next Model Released

Hugging Face · 5d ago

A high-download flash release from the Qwen series provides a fast and efficient option for local AI setups.

Models

High-End AI Models Face Adoption Hurdles Against Cheaper Tools

Simon Willison · 6d ago

Discussion highlights how expensive premier AI models are struggling to attract broad user bases while cheaper alternatives thrive.

Models

Why Your Local LLM Might Feel Dumber Than Expected

Hacker News · 7d ago

A community discussion explores common pitfalls, configuration issues, and prompt habits that make local language models underperform.

Models

Privacy Concerns Grow as Meta AI Glasses Gain Popularity

Ars Technica · 8d ago

As demand for smart glasses rises, users face challenges avoiding uninvited recordings, and detection apps remain imperfect.

General

Up to 3.2x faster inference with LFM2.5-DSpark

Hugging Face Blog · 9d ago

The new LFM2.5-DSpark architecture achieves up to 3.2x faster inference speeds for local AI workloads.

General

Grok data exfiltration via encrypted malicious instructions

Ars Technica · 9d ago

Researchers demonstrated that encrypting malicious instructions can bypass AI safety guardrails, causing Grok to exfiltrate user data.

Models

The Mojo programming language is now open source

Simon Willison · 11d ago

The Mojo programming language, often used in AI development, has officially been made open source.

General

Popular Qwen 3.8 27B Uncensored GGUF release

Hugging Face · 12d ago

An uncensored version of Qwen 3.8 27B packaged in GGUF format, a compressed file format for running models locally efficiently.

Models

First impressions of the Qwen 3.8 27B model

Simon Willison · 13d ago

A hands-on review noting that Qwen 3.8 27B is an excellent local model, though it tends to overthink tasks by default.

Models

Shrunk DeepSeek V4 Flash runs locally on Mac

Hacker News · 13d ago

A developer successfully shrank the DeepSeek V4 Flash model down to roughly 57GB so it could run locally and write code on a Mac.

Models

KoboldCpp v1.119 adds video generation support

GitHub — KoboldCpp · 13d ago

KoboldCpp v1.119 adds support for video generation and Image-to-Video capabilities using Minimax H3, requiring four specific model files.

Tools & Updates