🧠 Large Language Models

News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.

1,682 articles

Researchers Warn AI Workflows Can Be Hijacked to Leak Data Without Jailbreaking

Security researchers say enterprise AI workflows can be exploited to steal sensitive data even without prompt injection, account takeovers, or jailbreaking the underlying model. The vulnerability reportedly stems from how AI systems handle privileged access within automated pipelines. Source: cybersecuritynews.com

BIT Computer Partners with AWS to Build LLM for Korea's Healthcare Sector

South Korea's BIT Computer announced a partnership with Amazon Web Services to develop a large language model tailored for the healthcare industry. The companies aim to bring AI-driven tools into Korean medical services. Source: koreabiomed.com

Meitu Xiuxiu Rolls Out New AI Assistant Powered by Its Own Large Language Model

Chinese photo-editing app Meitu Xiuxiu has introduced a new AI Assistant feature built on its in-house large language model, according to Sina Technology. The update was announced on September 10, expanding the app's AI capabilities for users. Source: news.futunn.com

Mercury 2.5 diffusion model claims 1,107 tokens/sec at GPT-5.6 Luna-level quality

Inception unveiled Mercury 2.5 on September 8, 2026, a text-generation model built on diffusion technology rather than the typical autoregressive approach. The company says it matches GPT-5.6 Luna's performance while generating up to 1,107 tokens per second at a cost of roughly 120 yen per million output tokens. Source: gigazine.net

GPT-6 Astra meets V2Fun: large models begin orchestrating 3D generation

A new wave of AI tools is combining large language models with 3D generation pipelines, letting models coordinate multiple specialized systems instead of doing the work themselves. This shift points to a broader trend of AI acting as an orchestrator across code, image, and 3D modeling workflows. Source: eu.36kr.com

SCX.ai builds Swift package to tap Apple's on-device LLMs, with Australia in focus

SCX.ai is developing an MIT-licensed Swift package that lets developers connect to Apple's hosted large language models more flexibly. The project has a particular focus on Australian use cases and aims to simplify access to Apple's LLM capabilities. Source: arnnet.com.au

Apple's Siri AI debuts in English on September 14, more languages coming in October

Apple confirmed that its overhauled, LLM-powered Siri AI will launch alongside iOS 27 on September 14, initially supporting only English. Five additional languages are set to follow in October as the rollout expands. Source: gagadget.com

Another Anthropic employee quits, warns about frontier AI labs' direction

An Anthropic staffer known as Cal has publicly announced leaving the company, citing concerns about the path frontier AI labs are taking. The departure adds to a growing list of insiders voicing doubts about how leading AI companies are handling model development. Source: nationalreview.com

Amazon Study Uses Ising Models to Expose Bias in LLM Judges

New research from Amazon suggests that apparent agreement among LLM-based judges can be deceptive. Using Ising models borrowed from physics, researchers Krishna Balasubramanian and Sasha found hidden bias patterns behind seemingly strong consensus. Source: quantumzeitgeist.com

AI Model Prices Plunge Faster Than PCs Ever Did

The cost of running AI models has dropped dramatically over just three years, outpacing even the steep price declines seen during the personal computer boom of the 1980s. Analysts point to this rapid cost collapse as a key driver of wider AI adoption. Source: citybiz.co

South Korea to Build Medical LLM Trained on Local Clinical Standards

South Korea plans to develop a large language model specifically adapted to its domestic healthcare system. The move follows criticism that foreign AI models fail to properly reflect local medical practices and guidelines. Source: mbiz.heraldcorp.com

Inception Debuts Mercury 2.5, Claiming 1,100 Tokens per Second

Inception has released Mercury 2.5, a diffusion-based LLM said to hit over 1,100 tokens per second in real-world use. The launch highlights how speed is becoming a new front in the competition among AI model makers. Source: shattered.io

What Chaos Theory Reveals About Large Language Models

LLMs may look like black boxes, but researchers argue the math behind their behavior is well understood. A new deep-dive explores how chaos and complexity theory can explain patterns inside these models. Source: hackernoon.com

Cipheras Group's Apex AI Fund rolls out proprietary 290-billion-parameter LLM

A private AI-focused investment fund says it has deployed a custom-built 290-billion-parameter large language model. The move positions it among the first investment vehicles to use such a large proprietary model in its operations. Source: lincolnjournal.com

Grok 4.5 debuts with emphasis on coding and AI agents, heating up rivalry with OpenAI and Anthropic

Elon Musk's AI company xAI has released Grok 4.5, the newest version of its flagship large language model. The update targets coding capabilities and agentic AI features, pushing further competition with OpenAI and Anthropic. Source: ibtimes.com

Windows PCs aim to run 100-billion-parameter LLMs locally, challenging Mac's AI edge

After two years of hype around AI PCs, a new effort claims to move local execution of massive language models from demos into real, usable products on Windows machines. The goal is to close the gap with Apple's Mac lineup in on-device AI performance. Source: eu.36kr.com

Study finds language models causally rely on confidence to guide behavior

New research explores metacognition in AI systems, showing that confidence signals extracted from language models can causally influence their outputs. The findings suggest models don't just simulate self-assessment but actually use it to shape decisions, similar to biological cognition. Source: nature.com

Google Cloud says smaller Gemma 3 12B outperforms 27B model on TPU throughput

Google Cloud reports that under heavy load, the smaller Gemma 3 12B model kept scaling efficiently on TPU v6e hardware. In contrast, the larger 27B version hit a throughput ceiling once user load passed 64. Source: itbrief.asia

Why history makes me skeptical of Anthropic's potentially massive upcoming IPO

The AI startup Anthropic is reportedly preparing to file for a public listing after Labor Day, seeking a valuation near $2 trillion. The author outlines historical reasons for staying cautious about participating in the offering. Source: theglobeandmail.com

Questions raised over whether OpenAI's new model behaved unexpectedly

The article revisits a recent security incident where Hugging Face reported a breach of its production infrastructure. It raises the question of whether an OpenAI model played an unintended role in the episode. Source: predictiveanalyticsworld.com