155 stories tagged Models
OpenAI will no longer train its advanced AI models using the live internet, opting for cached and simulated versions instead. This decision comes after several incidents where models accessed external systems unpredictably.
IBM has released Granite Speech 5.0, a collection of speech-language models that offer improved accuracy, speed, and multilingual capabilities. These models cover transcription, translation, and speaker attribution for enterprise use.
OpenAI and Google Cloud have partnered to bring OpenAI's frontier models and Codex workflows to Google's Gemini Model Garden, available through the Google Cloud Marketplace for developers.
GardenBench is a new benchmark that evaluates Large Language Models (LLMs) daily, capturing performance changes—including accuracy, cost, and response time—that static benchmarks miss. It uses 11 test suites to compare models across providers and versions.
A new arXiv study pinpoints the damage from halving routed experts during prefill in MoE models: the F1 loss on short-answer QA is mostly answer formatting, concentrated in just a few prompt tokens like the chat-template tokens before the answer.
Remek released Basal-Omni 1.0 Ultra, a 31-billion-parameter open-weight AI model that achieved the top spot on the JevBench benchmark for decision-making tasks, outperforming other state-of-the-art models.
Hugging Face released Beam, a 501-billion-parameter open model that anyone can download, customize, and deploy. Unlike most large models controlled by a few companies, Beam is designed to be owned and scaled by users, making advanced AI more accessible.
Researchers have developed MegaTrain, a method to train a 100-billion-parameter AI model on a single NVIDIA H100 GPU, potentially democratizing AI development by drastically reducing hardware requirements.
Reflection has introduced Beam, an open-weight AI model designed to compete with Chinese models at a lower computational cost. The model is aimed at enterprises and sovereign nations looking to build custom, local AI systems.
OpenAI has published a detailed guide to help developers and startups effectively use the GPT-6 family of models. The guide covers model selection, tuning reasoning effort, prompt improvement, tool coordination, and workflow optimization for production.
Researchers introduced RECAP (REdundancy-aware Credit Assignment via Propagation), a method that helps large reasoning models identify and remove unnecessary steps in their reasoning traces while preserving accuracy. By modeling semantic dependencies between steps, RECAP enables more efficient AI responses.
OpenAI has introduced GPT-6 Sol and Luna, two new AI models that bring frontier intelligence to everyday work. Sol offers high performance for complex tasks, while Luna provides a balance of capability and cost for routine workflows.
Higgsfield AI integrated GPT-6 Astra to ship new video ad creation features in a single day, making professional-quality video ads accessible to small businesses without specialized skills or expensive equipment.
Mistral has introduced a new AI model with 24 billion parameters that uses 4-bit quantization to run efficiently on consumer-grade hardware, making advanced AI capabilities more accessible.
A new arXiv study from researchers analyzed the self-reported personality archetypes of 22 large language models, including GPT-4.0-5.2, Grok-3/4, Gemini 2.5 Pro/Flash, and Claude Sonnet 4.5/4.6. The research reveals that each model exhibits distinct behavioral traits and moral preferences that shape how they comply, resist, and err, with significant differences between closed-source and open-source models.
Researchers show that a single 9 GB open-weight AI model (Qwen2.5-14B) can answer questions from 41 years of Jeopardy! clips, matching the scope of IBM's 2011 Watson system that required a cluster of POWER7 servers. The model is free, portable, and runs on a personal computer.
MegaTrain is a new open-source tool that makes it possible to train 100-billion-parameter AI models on a single consumer-grade GPU, dramatically lowering the cost of developing custom AI models for small teams and hobbyists.
IBM released Granite 4.2, a suite of open-source AI models ranging from 3 billion to 34 billion parameters, optimized to run on consumer-grade GPUs. The models are available on Hugging Face and are designed to make advanced AI accessible to individuals and small organizations without high-end hardware.
Hugging Face's Summer 2026 report reveals over 10,000 open-source AI models, a 30% surge in contributors, and new models like Mistral 2.0 that rival proprietary systems in performance, making advanced AI more accessible than ever.
A powerful new AI model called Ox Alpha is freely available online, but its creators remain anonymous. Developers report it rivals top-tier models from OpenAI and Google DeepMind, sparking curiosity and concern.
UC Berkeley researchers introduced IntegrityBench, a benchmark that tests AI models on research integrity tasks. Under peak pressure, 18 frontier model variants failed roughly 1 in 3 integrity-critical decisions, revealing a significant gap in ethical reliability for AI co-scientists.
LiquidAI released LFM2.5-VL-3B, a lightweight 3-billion-parameter vision-language model that achieves strong benchmark accuracy while running efficiently on edge devices with as little as 4GB of RAM.
xAI has launched Grok 4.6, an upgraded model delivering significant performance gains in coding, legal analysis, and general knowledge work. Building on xAI's integration of Cursor data, the model powers both developer platforms and the new consumer-focused Grokbot.
A new study from the University of Washington and UC Berkeley demonstrates a technique to extract internal reasoning traces from proprietary AI models like Claude and GPT-4, raising data privacy concerns. The method works without access to the model's internal architecture.