Morning, {{first name | folks}}! NVIDIA is claiming a 5x throughput boost for AI agents, Anthropic is letting outside researchers examine its safety work, and OpenAI is opening its life sciences model to university labs. Plus, Meta is pushing a massive model toward local hardware, and CERN is bringing Mistral into particle physics. Let’s get into it.
Today’s Top 5
NVIDIA’s New Open Model Targets Faster AI Agents: NVIDIA says its 120B-parameter model delivers up to 5x higher throughput on agentic tasks than comparable open models.
Anthropic Opens Its AI Safety Work to Six Outside Researchers: Six METR researchers can examine Anthropic’s internal safety work and publish their findings without company approval.
OpenAI Offers University Labs Free Access to Its Life Sciences AI: Eligible U.S. university labs can get up to 30 free seats for 12 months to use OpenAI’s model for life sciences research.
Meta Releases an 816B AI Model for Local Use: Meta’s Muse Spark Open targets coding and agents, and the company says it can run on two Apple M5 Ultra Mac Studios.
CERN Turns to Mistral to Analyse Particle Collisions: CERN and Mistral are developing an AI model to reconstruct particle collisions and identify unusual signals as future experiments generate far more data.
NVIDIA has introduced Nemotron 3.5 Super VL, an open 120B-parameter model built for AI agents that reason, use tools and process text, images, documents and video. NVIDIA claims up to 5x higher throughput than comparable open models on agentic tasks.
Companies are also adapting it for specialized work. NVIDIA reports 40x lower inference costs for automated vendor-credit processing and a 41% accuracy improvement in threat analysis, highlighting the potential of models tailored to specific business workflows.
Anthropic has given six researchers from nonprofit METR access to its internal tools, systems and risk assessment teams to examine its AI models and safety practices. The team will be able to publish its findings without Anthropic’s approval, although the company can require limited redactions to protect sensitive information.
The move gives outside researchers a closer look at how Anthropic assesses its models, rather than relying only on public reports. The company is also working with SecureBio and Accenture on separate evaluation efforts.
OpenAI has launched the Rosalind Academic Plan, offering eligible U.S. university research labs up to 30 free seats for 12 months. The program gives academic teams access to ChatGPT Business and Codex, helping them work with research data, review literature, and develop analysis code.
The plan also includes access to GPT-Rosalind, OpenAI’s specialized life sciences model, subject to approval. Built for research across drug discovery, genomics and medicinal chemistry, it gives academic labs a way to explore specialized AI tools without paying for the seats during the sponsored period.
Meta has released Muse Spark Open, an 816-billion-parameter model aimed at developers building coding tools and AI agents. It activates 42 billion parameters per token and supports a one-million-token context window, giving it room to handle large codebases and lengthy tasks.
The model can run on two Apple M5 Ultra Mac Studios, according to Meta, and its weights are available for developers to download and customize. It’s an unusual desktop setup for a model this large, offering developers another option beyond cloud-based access.
CERN is teaming up with Mistral AI to build a model that can help scientists analyse particle collisions from the Large Hadron Collider. The system will be trained on simulated collision data to help identify particles, reconstruct events and spot unusual signals.
The timing matters: CERN expects its planned collider upgrade around 2030 to produce roughly ten times more collisions. Researchers need faster ways to make sense of that data, and the partnership aims to bring several analysis tasks into one AI model.
Other AI Signals:
ByteDance has created an AI Alignment & Preparedness team and plans to hire 10 researchers globally to focus on model safety, risk detection, and alignment.
MIT researchers have developed an AI approach to decode how cells communicate as they form tissues, with potential applications in stem-cell engineering and disease research.
Google is updating AI Search with more prominent source links and new ways to discover original websites through AI Overviews and AI Mode.
AWS says its security agent successfully exploited 33 of 37 real-world vulnerabilities in a test of autonomous penetration-testing capabilities.
IBM has released Granite Switch 4.1, with 8B and 30B models that handle language, decision-making, safety checks, and speech-to-text through one hosted endpoint.
AI Tools to Try:
Reason: A coding agent designed to get more done with fewer tokens.
Spinachi: Organizes health records, lab reports and reminders through WhatsApp.
Fluid Functionalism: Open-source UI components that use motion to make interactions easier to understand.
Hano: Creates and animates polished 3D device mockups directly in your browser.
