News about AI models

NVIDIA released its open-weight Nemotron 3.5 Super VL, a 120-billion-parameter model, on October 15, and on October 19 the German government and NVIDIA signed a Joint Declaration of Intent to formalize an industrial AI ecosystem. At GTC Berlin on October 21, the company introduced “sovereign intelligence” and DSX PODs that can cut onsite deployment from six months to 14 days. The article argues that open models spread NVIDIA's software, while revenue comes from the compute infrastructure needed to run them. It also corrects several reports: the Nemotron Coalition began in March 2026 rather than launching in Berlin, the roughly $2 billion figure was media expectation, and NVIDIA says it has committed nearly $10 billion across Europe.

Multiverse Computing launched Comet 60B, an open-source multimodal model built by compressing and optimizing NVIDIA's Nemotron 3.5 Super VL, which has 120 billion total parameters. Comet 60B halves total parameters to 60 billion while keeping 12 billion active parameters and support for text, image and video, so it fits on less hardware. The company says it stays within 5% of the original's accuracy and improves on some benchmarks, and it scores above the previous-generation Nemotron 3 Super on all eight benchmarks shown. The model was built with the CompactifAI framework through tensor-network profiling, compression, healing and GRPO-style reinforcement learning, and ships in BF16, FP8 and NVFP4 formats. Multiverse says it is aimed at developers building agentic AI workloads.

Nvidia (NVDA) is expanding its open AI model lineup with three new Nemotron models: Nemotron 3.5 Super VL, Nemotron 3.5 Bolt 74B-A9B 24GB, and Nemotron Bolt E4B 4GB, built to run agents from data centers to desktops. Super VL targets tasks lasting days or weeks and can be post-trained on enterprise data, Bolt 74B handles general agentic tasks like browsing and app use, and Bolt E4B is aimed at physical and conversational agents. The company argues open-weight models are cheaper and more private, since companies can run them on their own hardware, which could drive more chip sales over time. China currently leads the open-model space with models like Kimi K3 and Qwen, and CEO Jensen Huang has urged US firms to build more open models to boost competition.

Meta has launched Muse Wonder, a frontier AI model designed for personal and professional AI agents, comprising the full Muse Wonderful model and a faster version. It will first roll out to some subscribers and eventually power Muse, Meta's AI agent. CEO Mark Zuckerberg said in a post on X that Meta has caught up with OpenAI and Anthropic, marking its return to the frontier after more than a year on the sidelines, following the scrapped Llama 4 Behemoth and a $14 billion acquihire of Scale AI co-founder Alexandr Wang, now Chief AI Officer. The discussion notes that these claims rest on benchmarks and that real-world developer feedback will be the true test.