Ad
Short

Hugging Face has launched HuggingChat Omni, an AI router that selects the best open source model for each user prompt from more than 100 available models. The system automatically chooses the fastest, cheapest, or most suitable model for each task, using an approach similar to the new GPT-5 router. Supported models include gpt-oss, qwen, deepseek, kimi, and smolLM.

Image: Hugging Face

Hugging Face co-founder Clément Delangue says HuggingChat Omni is just the beginning. The platform already offers access to more than two million open models, spanning not only text, but also images, audio, video, biology, chemistry, time series, and more.

The routing system is built on Arch-Router-1.5B from Katanemo, a lightweight 1.5 billion-parameter model that classifies queries by topic and action. Arch-Router claims it outperforms other models at matching human preferences and is fully open source. Details are available in the research paper on arXiv.

Short

OpenAI has paused the use of Martin Luther King Jr.’s likeness in its Sora video tool after users created offensive content, including a clip showing him making monkey noises at a lectern. The King estate called for the ban.

OpenAI says it will revise its rules for deepfakescalled "cameos" in Sora. Although the company cites "strong free speech interests in depicting historical figures," it now says families and authorized groups should control how public figures are portrayed. Authorized representatives or estate owners can request removal from the tool.

Short

With version 1.5, Manus is introducing its most capable AI agent system so far. The updated architecture cuts task processing times from 15 minutes to just 4, and internal tests report a 15 percent jump in output quality. Manus claims the agent can now build, test, and refine full web applications—including backend, user management, and database—directly on the platform. It also handles research, image generation, and analysis.

"We didn’t create a 'website builder' feature. We enhanced the core Manus platform to master a new, complex domain," Manus co-founder Tao Zhang says.

Both Manus-1.5 and Manus-1.5-Lite are launching as new releases. The Lite version is open to everyone, while the full version is reserved for subscribers. Microsoft has also integrated Manus into its agent suite for Windows.

Ad
Ad
Short

Anthropic's Claude now integrates directly with Microsoft 365, letting users pull in content from Outlook, SharePoint, OneDrive, and Teams. This means emails, calendars, and documents can show up right in a Claude chat, making it easier to bring company data into the conversation.

Anthropic is also adding a centralized search that covers company resources like HR documents and team guidelines. These features are available to Claude Team and Enterprise users, but admins have to enable them first.

The rollout follows Microsoft's recent Copilot integration in Windows, showing just how closely the competition for AI in the workplace is tied to the digital office ecosystem.

Ad
Ad
Short

Google Labs Creative Director Henry Daubrez says the new Veo 3.1 update is being overhyped. Though it adds helpful features like image-to-image animation, he sees the changes as minor. Daubrez blames financial pressure in the AI industry for turning small updates into big marketing moments. Veo is part of Flow by Google.

"The bigger issue is that the enormous financial stakes around AI are turning timelines into marketing noise, with rumors inflated to the point where every incremental update gets treated like a paradigm shift."

Henry Daubrez

Short

Google Deepmind and Yale University have introduced a new AI model called C2S-Scale 27B, built on the open Gemma model family. The system analyzes single cells and, according to Google Deepmind, uncovered a previously unknown therapeutic pathway for cancer. The model identified the drug silmitasertib (CX-4945) as a "conditional enhancer" - meaning it can make tumor cells more visible to the immune system under specific conditions. In the team's words:

"This result also provides a blueprint for a new kind of biological discovery. It demonstrates that by following the scaling laws and building larger models like C2S-Scale 27B, we can create predictive models of cellular behavior that are powerful enough to run high-throughput virtual screens, discover context-conditioned biology, and generate biologically-grounded hypotheses."

Google Deepmind

Lab experiments with human neuroendocrine cell models confirmed the prediction. Before that, C2S-Scale simulated over 4,000 drugs in two immune environments. The code is available on GitHub, and the model can be found on Hugging Face. More details are in the preprint on bioRxiv.

Ad
Ad
Short

Apple's new M5 chip is aimed at speeding up AI features on the MacBook Pro, iPad Pro, and Apple Vision Pro. Apple says the M5 delivers more than four times the GPU performance for AI tasks compared to the previous M4 chip. Each of the ten GPU cores includes a built-in "Neural Accelerator." The 16-core Neural Engine can process up to 38 trillion operations per second, making it 60 percent faster than the Neural Engine in the M4. The CPU is also up to 15 percent quicker at multi-threaded processing.

The chip uses a third-generation 3 nm manufacturing process, offers 153 GB/s of shared memory bandwidth (about 30 percent more than before), and supports up to 32 GB of unified memory. This allows larger AI models to run directly on the device. Apple expects that existing AI apps and tools like Apple Intelligence will see faster performance. Developers can access the AI hardware using Core ML, Metal 4, and Tensor APIs.

Google News