NVIDIA pairs European agentic builds with Nemotron 3.5 Super VL

NVIDIA detailed French and Nordic agentic AI deployments and released open 120B Nemotron 3.5 Super VL for enterprise agents.

NVIDIA’s Oct. 13 blog says French companies are deploying agentic computing on its stack: Bull plans European production of Vera Rubin Ultra NVL72 systems for 2028, and the Altair-1 satellite, integrated by Loft Orbital, runs Jetson AGX Orin modules and Mistral models. LINAGORA and Docaposte’s CANUT agreement makes Nemotron-based Twake.ai the referenced sovereign workplace for more than 2,500 public organizations, and Mistral placed open-weight Mistral Large 4 into private preview. An Oct. 15 NVIDIA blog treats the Nordics separately, projecting roughly 1 gigawatt of AI-cloud capacity by the end of 2027 across CoreWeave, Crusoe, and Nscale, and Sweden’s Sferical One on two DGX SuperPODs with 16 GB300 racks and 1,152 GPUs, with Spotify, ABB, Klang, and Einride among the users cited. The Decoder and MarkTechPost both report the same day’s release of open Nemotron 3.5 Super VL, a 120-billion-parameter mixture-of-experts model with about 12 billion parameters active per token, distilled from Nemotron 3 Ultra for text, images, documents, and video. They do not contradict shared size figures; The Decoder adds coalition partners, CharXiv at 80.8 percent (84 percent with Python), Dynamo gains on 16 B200 GPUs, and a later 4GB Bolt E4B, while MarkTechPost adds a 1 million-token context, OpenMDW-1.1 weights, about 26 trillion training tokens, video up to two minutes, NVFP4 scores including 71.07 on SWE-bench Verified and 89.90 on GPQA-Diamond, and says edge Bolt models shipped with the release.

• Bull plans European production of NVIDIA Vera Rubin Ultra NVL72 systems for 2028; Altair-1, integrated by Loft Orbital, runs Jetson AGX Orin modules and Mistral models.
• The CANUT agreement makes Nemotron-based Twake.ai the referenced sovereign workplace for more than 2,500 French public organizations; open-weight Mistral Large 4 is in private preview.
• Nordic AI-cloud capacity is projected at about 1 GW by the end of 2027 across CoreWeave, Crusoe, and Nscale; Sferical One uses two DGX SuperPODs, 16 GB300 racks, and 1,152 GPUs.
• ABB cites up to 99% simulation accuracy; Klang’s Pianissimo cuts Swedish transcription errors by 75% and is 64x faster; Einride is building freight trucks on NVIDIA Hyperion.
• Nemotron 3.5 Super VL is an open 120B MoE multimodal model with about 12B parameters active per token, distilled from Nemotron 3 Ultra; The Decoder says that is about one-fifth Ultra’s size.
• The Decoder reports CharXiv at 80.8% (84% with Python) and, with Dynamo on 16 B200 GPUs, eight times as many concurrent users and an 85% lower time to first token; Bolt 74B-A9B targets 24GB, with a 4GB Bolt E4B planned later.
• MarkTechPost reports a 1M-token context, OpenMDW-1.1 weights, video up to two minutes, training on about 26 trillion tokens, and NVFP4 scores of 89.90 on GPQA-Diamond, 71.07 on SWE-bench Verified, and 77.5 on BrowseComp.