Key Technical Specifications

• Architecture: Hybrid Mamba-Transformer Mixture-of-Experts (MoE), combining the long-context efficiency of state-space models with the reasoning power of transformers .
• Parameters: 550B total, with ~55B active per token .
• Context Window: A native 1 million tokens, enabling the model to reason over the equivalent of several long novels or massive enterprise codebases in a single pass .
• Key Innovation 1: LatentMoE. This technique compresses token representations into a low-rank space before routing them to experts. NVIDIA says this allows for 4× as many expert specialists for the same inference cost, boosting accuracy per FLOP .
• Key Innovation 2: Multi-Token Prediction (MTP). The model predicts multiple future tokens simultaneously in a single forward pass, which improves chain-of-thought coherence and acts as a form of built-in speculative decoding for faster text generation .
• Training Precision: Pre-trained using NVIDIA's ultra-efficient 4-bit NVFP4 format on the Blackwell architecture .
• Performance: NVIDIA claims 5x faster inference compared to the prior generation of models .

These specifications are purpose-built for "agentic AI." A 1M-token context window and native multi-token prediction are not designed for casual chat; they are designed for a model that must maintain an internal memory of a complex workflow, reason about code, monitor a data pipeline, and plan multistep actions over an extended period . NVIDIA has positioned the Ultra variant specifically as a "large reasoning engine for complex AI applications" requiring deep research and strategic planning . The open-source weights for the model were scheduled for public release on June 4, 2026, via Hugging Face, OpenRouter, and other platforms .

Palantir's Agentic Platform: AI FDE Meets the Ontology

Palantir is not simply plugging a chatbot into its system. It is integrating Nemotron models into its AI FDE (Forward Deployed Engineer) platform, an interactive agent that operates within Palantir Foundry . AI FDE is designed to translate natural language commands into concrete operational actions: performing data transformations, managing code repositories, and building and maintaining an organization's central ontology .

The integration aims to make these agents "long-running," a term NVIDIA and Palantir are using to distinguish this new wave from single-turn LLM queries . A long-running agent in this context can autonomously execute a complex, multi-step task—such as ingesting a new data stream, transforming it, updating the ontology, and building a new operational application on top of it—without human hand-holding at every step . The agent is designed to learn continuously from these interactions, enabling the construction of domain-specific, air-gapped enterprise systems that deepen their specialization over time .

This capability is anchored by Palantir's core differentiator: its Ontology. The Ontology is a digital representation that maps all of an organization's data, logic, and actions, allowing an AI agent to understand not just the data, but how the business actually works . By fusing NVIDIA's Nemotron models with this semantic map, the two companies are building a stack designed for real-time operational decision-making, where the AI understands the ripple effect of an action across a supply chain, a military logistics network, or a cybersecurity posture .

The Customer Base: Enterprise and Government

This technology is not hypothetical. It is being targeted at some of the world's most complex and mission-critical environments. Palantir's customer base spans two broad but deeply intertwined sectors:

• Enterprise: Early adopters of the broader NVIDIA-Palantir AI infrastructure include major commercial players like Lowe's, Booz Allen Hamilton, and T-Mobile . The platform is designed for operational decision-making across supply chains, retail, healthcare, and financial services .
• Government and Defense: Through its Gotham and AIP platforms, Palantir serves a vast defense and intelligence apparatus. The company has forged strategic partnerships to accelerate this AI push, including a collaboration with Accenture Federal Services announced in June 2025 to deploy AI for federal mission-critical workflows, and a 2026 alliance with Deloitte to build an Enterprise Operating System for unlocking data trapped in government agency silos .

The end vision, demonstrated at Palantir's DevCon conferences, is an AI agent that can manage the entire engineering lifecycle within a secure environment—from writing functions and authoring evaluations to safely debugging code in a branch-aware loop—all within systems that are often completely disconnected from the public internet .

A Strategy Months in the Making

Monday's announcement is a key product milestone, not the start of the relationship. The strategic framework for this integration was laid at GTC Washington D.C. in October 2025, when NVIDIA and Palantir first announced they were building a "first-of-its-kind integrated technology stack for operational AI" . That initial pact committed to combining NVIDIA's Blackwell architecture, CUDA-X libraries, and Nemotron models with Palantir's Ontology platform .

Alongside the Palantir news, NVIDIA's GTC Taipei event on June 1 served as a broader launchpad for its enterprise agent strategy. The company introduced the NVIDIA Agent Toolkit, a platform bundling its NemoClaw blueprints, Nemotron models, OpenShell secure runtime, and CUDA-X libraries to help enterprises deploy autonomous AI agents. Palantir and SAP were named as the marquee launch partners . Cybersecurity giant CrowdStrike also announced a parallel integration, using Nemotron models to power new vulnerability identification agents, confirming that the "long-running agent" paradigm is an ecosystem-wide push, not a one-off partnership .

What It Means

The Palantir-NVIDIA integration signals a maturation of enterprise AI from experimental copilots to embedded, autonomous operators. By pairing a state-of-the-art open model optimized for long-horizon reasoning with a platform that already represents how an organization runs, the two companies are betting that the AI agent's natural habitat isn't a chat window—it's the operational guts of the business itself. The open-source nature of Nemotron 3 Ultra also gives security-conscious government and enterprise customers a path to deploy cutting-edge reasoning on private, air-gapped infrastructure without sending data to third-party APIs, a non-negotiable requirement for Palantir's core market .

Key Technical Specifications

• Architecture: Hybrid Mamba-Transformer Mixture-of-Experts (MoE), combining the long-context efficiency of state-space models with the reasoning power of transformers .
• Parameters: 550B total, with ~55B active per token .
• Context Window: A native 1 million tokens, enabling the model to reason over the equivalent of several long novels or massive enterprise codebases in a single pass .
• Key Innovation 1: LatentMoE. This technique compresses token representations into a low-rank space before routing them to experts. NVIDIA says this allows for 4× as many expert specialists for the same inference cost, boosting accuracy per FLOP .
• Key Innovation 2: Multi-Token Prediction (MTP). The model predicts multiple future tokens simultaneously in a single forward pass, which improves chain-of-thought coherence and acts as a form of built-in speculative decoding for faster text generation .
• Training Precision: Pre-trained using NVIDIA's ultra-efficient 4-bit NVFP4 format on the Blackwell architecture .
• Performance: NVIDIA claims 5x faster inference compared to the prior generation of models .

These specifications are purpose-built for "agentic AI." A 1M-token context window and native multi-token prediction are not designed for casual chat; they are designed for a model that must maintain an internal memory of a complex workflow, reason about code, monitor a data pipeline, and plan multistep actions over an extended period . NVIDIA has positioned the Ultra variant specifically as a "large reasoning engine for complex AI applications" requiring deep research and strategic planning . The open-source weights for the model were scheduled for public release on June 4, 2026, via Hugging Face, OpenRouter, and other platforms .

Palantir's Agentic Platform: AI FDE Meets the Ontology

Palantir is not simply plugging a chatbot into its system. It is integrating Nemotron models into its AI FDE (Forward Deployed Engineer) platform, an interactive agent that operates within Palantir Foundry . AI FDE is designed to translate natural language commands into concrete operational actions: performing data transformations, managing code repositories, and building and maintaining an organization's central ontology .

The integration aims to make these agents "long-running," a term NVIDIA and Palantir are using to distinguish this new wave from single-turn LLM queries . A long-running agent in this context can autonomously execute a complex, multi-step task—such as ingesting a new data stream, transforming it, updating the ontology, and building a new operational application on top of it—without human hand-holding at every step . The agent is designed to learn continuously from these interactions, enabling the construction of domain-specific, air-gapped enterprise systems that deepen their specialization over time .

This capability is anchored by Palantir's core differentiator: its Ontology. The Ontology is a digital representation that maps all of an organization's data, logic, and actions, allowing an AI agent to understand not just the data, but how the business actually works . By fusing NVIDIA's Nemotron models with this semantic map, the two companies are building a stack designed for real-time operational decision-making, where the AI understands the ripple effect of an action across a supply chain, a military logistics network, or a cybersecurity posture .

The Customer Base: Enterprise and Government

This technology is not hypothetical. It is being targeted at some of the world's most complex and mission-critical environments. Palantir's customer base spans two broad but deeply intertwined sectors:

• Enterprise: Early adopters of the broader NVIDIA-Palantir AI infrastructure include major commercial players like Lowe's, Booz Allen Hamilton, and T-Mobile . The platform is designed for operational decision-making across supply chains, retail, healthcare, and financial services .
• Government and Defense: Through its Gotham and AIP platforms, Palantir serves a vast defense and intelligence apparatus. The company has forged strategic partnerships to accelerate this AI push, including a collaboration with Accenture Federal Services announced in June 2025 to deploy AI for federal mission-critical workflows, and a 2026 alliance with Deloitte to build an Enterprise Operating System for unlocking data trapped in government agency silos .

The end vision, demonstrated at Palantir's DevCon conferences, is an AI agent that can manage the entire engineering lifecycle within a secure environment—from writing functions and authoring evaluations to safely debugging code in a branch-aware loop—all within systems that are often completely disconnected from the public internet .

A Strategy Months in the Making

Monday's announcement is a key product milestone, not the start of the relationship. The strategic framework for this integration was laid at GTC Washington D.C. in October 2025, when NVIDIA and Palantir first announced they were building a "first-of-its-kind integrated technology stack for operational AI" . That initial pact committed to combining NVIDIA's Blackwell architecture, CUDA-X libraries, and Nemotron models with Palantir's Ontology platform .

Alongside the Palantir news, NVIDIA's GTC Taipei event on June 1 served as a broader launchpad for its enterprise agent strategy. The company introduced the NVIDIA Agent Toolkit, a platform bundling its NemoClaw blueprints, Nemotron models, OpenShell secure runtime, and CUDA-X libraries to help enterprises deploy autonomous AI agents. Palantir and SAP were named as the marquee launch partners . Cybersecurity giant CrowdStrike also announced a parallel integration, using Nemotron models to power new vulnerability identification agents, confirming that the "long-running agent" paradigm is an ecosystem-wide push, not a one-off partnership .

What It Means

The Palantir-NVIDIA integration signals a maturation of enterprise AI from experimental copilots to embedded, autonomous operators. By pairing a state-of-the-art open model optimized for long-horizon reasoning with a platform that already represents how an organization runs, the two companies are betting that the AI agent's natural habitat isn't a chat window—it's the operational guts of the business itself. The open-source nature of Nemotron 3 Ultra also gives security-conscious government and enterprise customers a path to deploy cutting-edge reasoning on private, air-gapped infrastructure without sending data to third-party APIs, a non-negotiable requirement for Palantir's core market .