Nvidia launches open models to run intelligent agents inside data centers and personal devices
Nvidia moved to encircle its customers and competitors at the same time, unveiling a suite of open-weight models designed to run intelligent agents inside private data centers and directly on desktop computers. The step, announced by the company's vice-president of open-source models and software, Kari Praske, puts in the hands of companies the Nemotron 3.5 Super VL model, comprising 120 billion parameters, to serve as an engine for complex-task agents, without forcing organizations to choose between model efficiency and ease of deployment on their own infrastructure.
Competition no longer revolves solely around parameter count, but around execution speed and data control.Nvidia states that its new model, although smaller in parameter count than the previous Nemotron 3 Ultra, delivers comparable accuracy in instruction following and tool invocation, and outpaces advanced models such as Anthropic’s Opus 5, GPT-6.1 Sol, Luna and Astra from OpenAI in agent-task completion speed after subsequent fine-tuning. The specialized tasks include document intelligence, financial accounting, scientific experimentation, medical insurance claim management, and assessing system resilience to security attacks; these capabilities have already attracted organizations such as EY, Salesforce, CrowdStrike, Perplexity, Sky Dances and Ramp.
Behind this technical launch lies a deep commercial influence battle. Major labs that own closed models have begun actively seeking alternatives to Nvidia chips, turning to Amazon’s Tranium processors, Cerebras wafers, or building their own chips in-house. Nvidia responded by going directly to organizations that prefer to own what they build, betting on companies’ tendency to fine-tune models on their confidential data and run them locally, rather than channeling sensitive business flows through cloud interfaces owned by third parties.
The agent race is shifting from massive data centers to office computers directly.In parallel, the company launched two lightweight models in the Bolt series, designed to run locally on AI computers equipped with the RTX Spark chip and workstations powered by the DGX Spark chip, which are currently being offered in the market by companies such as Microsoft and Dell. The first model, Bolt 74 B, carries 74 billion parameters on 24 GB of memory, balancing response speed, multi-step reasoning, web browsing and tool invocation, while later this month the Bolt E 4 B model, aimed at conversational agents, targets users and researchers who refuse to grant cloud agents unrestricted access to their personal records and daily lives.
This shift reshuffles technology-spending priorities for business sectors in the Gulf, Egypt and the Arab region. Financial institutions, hospitals and government agencies that have heavily invested in AI computing infrastructure face strict regulations that prohibit sending customer data and account records through external cloud APIs. Providing an open model of 120 billion parameters enables engineering teams to train agents specialized in accounting and document processing within their closed sovereign environments, reducing recurring cloud-consumption bills and turning local hiring needs from merely writing prompts to engineering post-training and fine-tuning open-weight models on on-premise servers.
At AI Gossips we observed how cloud-computing labs try to impose recurring fees on every programmatic call to intelligent agents, but the release of these open weights gives a technology leader a clear option: invest your team’s capabilities in your own infrastructure, rather than tying daily business flows to volatile external subscriptions.