IBM has released four new open-source Granite 4.0 Nano language models, ranging from 350 million to 1.5 billion parameters. These models prioritize efficiency and accessibility, designed to run on laptops and edge devices rather than requiring extensive cloud computing resources. The smallest models can even operate within a web browser, making them highly versatile for developers. The models are licensed under Apache 2.0, enabling commercial use and modification. They possess native compatibility with tools like llama.cpp and vLLM and are certified under ISO 42001 for responsible AI development. Benchmarks reveal they rival or outperform larger models in similar categories, particularly in instruction following and function calling. IBM's Granite models address needs for deployment flexibility, inference privacy, and open auditability. IBM has engaged with the open-source community, hinting at larger models and fine-tuning recipes in the future. This release signals a shift towards strategically scaled AI rather than solely relying on model size. Granite models are designed as enterprise-ready systems emphasizing transparency and performance.

ChatGPT, Qwen, and DeepSeek are the three most popular AI models. We put them through their paces with a series of key challenges. The results show which model is the smartest choice for your needs (and budget)

The ARC-AGI-2 test, developed by the Arc Prize Foundation, reveals a stark gap between human reasoning and current AI models....

Mark Zuckerberg announced that Meta AI is releasing four new open-source Llama language models, with two available now and two more coming soon. The goal is to make AI universally accessible and benefit everyone in the world. The first model, Llama 4 Scout, is extremely fast and has an industry-leading 10M-token context length, making it the highest performing small model in its class. The second model, Llama 4 Maverick, beats other models on benchmarks and is smaller and more efficient than DeepSeek v3. Zuckerberg also teased Llama 4 Reasoning and Llama 4 Behemoth, with the latter having over 2 trillion parameters and already being the highest performing base model in the world. The models can be downloaded from llama.com and Hugging Face, and can be used in WhatsApp, Messenger, or Instagram Direct. Meta AI believes that openness drives innovation and is good for developers, the company, and the world. The release marks the beginning of a new era of natively multimodal AI innovation, with the potential to lead to better products and more opportunities for developers. The community is encouraged to build new experiences with the Llama 4 models.

OpenAI's Aardvark, a GPT-5 agent, automates code security like a human researcher. It embeds in the development pipeline, shifting security to a continuous process. Aardvark reasons about code, automates threat detection, and verifies vulnerabilities. It maps repositories, builds threat models, and monitors new code commits for risks. Crucially, Aardvark validates exploitability in a sandbox to reduce false positives. Upon confirmation, it uses Codex to propose and analyze patches. This fixes vulnerabilities and ensures no new issues are introduced. One benefit is the significant reduction of false positives. Benchmarks show Aardvark identifies 92% of vulnerabilities in test repositories. This illustrates AI's potential to aid in code auditing.

A discrepancy between first- and third-party benchmark results for OpenAI’s o3 AI model is raising questions about the company’s transparency and model testing practices. When OpenAI unveiled o3 in...

This paper proposes FieldWorkArena, a benchmark for agentic AI targeting real-world field work. With the recent increase in demand for agentic AI, they are required to monitor and report safety and health incidents, as well as manufacturing-related incidents, that may occur in real-world work envir…

This table evaluates the impact of multi-token prediction on Llama 2 fine-tuning, suggesting that it does not significantly improve performance on various tasks

Benchmark analyst Mark Palmer hiked his Hut 8 price target to $36 from $33, while reiterating his buy rating on the stock.

New reports indicate significant security vulnerabilities in the recently released GPT-5. Two independent firms have tested the model, and both found its security to be lacking. One firm successfully "jailbroke" GPT-5 within 24 hours using a combination of existing techniques and storytelling. This allowed the model to generate instructions for creating a Molotov cocktail. The researchers highlighted that this exploit demonstrates the difficulty AI models face in preventing context manipulation. They found that multi-turn attacks can bypass single-prompt filters by exploiting the full conversational context. Simultaneously, another security firm declared the raw GPT-5 model "nearly unusable for enterprise out of the box." Their red teamers found that obfuscation attacks, such as inserting hyphens between characters, remain effective. This firm also found that OpenAI's internal prompt layering leaves significant gaps in business alignment. Benchmarking against GPT-4o, they concluded that GPT-4o remains the more robust model, especially when hardened. Both findings strongly advise approaching the current GPT-5 with extreme caution.

Learn how to compare LLMs using your own interal benchmark