Dell Enterprise Hub Brings Frontier Open Models to PowerEdge XE9780

Key takeaways:

• • Dell Enterprise Hub is adding five next-generation open models optimized for Dell PowerEdge XE9780 with NVIDIA B300 GPUs, delivering frontier-level on-premises AI performance with better token economics, data sovereignty and architectural control.
  • The blog highlights a shift from “open source vs. proprietary” to AI systems where open and frontier models coexist in multi-agent, long-context workflows, giving enterprises transparency, customization with their own data, and more predictable costs.
  • Dell Enterprise Hub, combined with multi-platform optimization, built-in security, Goodput-based sizing and the Dell AI SDK, enables enterprises to deploy and manage these advanced models quickly and reliably across NVIDIA and AMD infrastructure.

The AI ecosystem is undergoing a fundamental shift toward efficiency, accessibility and cost optimization. Although token pricing has plummeted, total AI spending continues to rise as multi-agent and autonomous workflows consume tokens at massive scale, making token economics a critical planning metric. Consequently, the industry is adopting a “smarter and leaner” approach across models, pricing and infrastructure. Hardware innovation has moved beyond raw GPU power to focus on high-bandwidth memory, hardware-software co-optimization and full-stack orchestration platforms.

Reflecting this transition, today Dell Enterprise Hub announced the addition of five next-generation open-source models — Kimi K2.6, DeepSeek V4 Pro, GLM 5.1, MiniMax M2.7 and DeepSeek V4 Flash.

To support the massive context and concurrency requirements of these new trillion-parameter, NVFP4-optimized models, hardware like the Dell PowerEdge XE9780 and liquid-cooled XE9780L offers the ideal platform. Powered by 8x NVIDIA HGX B300 NVL8 GPUs — each featuring 288GB HBM3 memory — these servers utilize full NVIDIA NVLink interconnects delivering 900GB/s per GPU bandwidth to eliminate communication bottlenecks. Ultimately, deploying these state-of-the-art open-source models on Dell’s B300-equipped infrastructure delivers frontier-level performance at a reasonable cost, complete with data sovereignty and architectural control.

The open source and frontier model paradigm

Modern AI is no longer defined by a single model but by orchestrated systems that integrate numerous models, autonomous agents, diverse data sources and memory layers — where open-source and frontier models coexist, each serving specific enterprise needs. Our previous additions to Dell Enterprise Hub, including Kimi K2.5, Trinity Large, Cohere Transcribe, Gemma-4 and Mistral-Small-4 reinforces this strategic shift, moving the industry beyond the “Open Source vs. Proprietary” debate toward an ecosystem where both thrive together. Within these systems, open-source models play a critical role by offering enterprises full auditability for compliance and security, the freedom to fine-tune with proprietary data for unmatched specialization, predictable licensing costs that scale economically and access to a global research community that innovates faster than any single vendor.

Model analysis: Technical deep dive

Kimi K2.6 (Moonshot AI)

Architecture: 1T parameter Mixture-of-Experts (MoE) with 32B active parameters, 256K context window, MIT license.

NVFP4 Optimization: Kimi K2.6 leverages NVIDIA’s custom FP4 quantization (NVFP4), reducing memory footprint by 75% while maintaining accuracy. This enables deployment of trillion-parameter models on GPU configurations without sacrificing performance and delivering bleeding-edge reasoning at unprecedented inference speeds on NVIDIA B300 GPUs.

Technical Differentiation: Native multi-agent swarm capability enables coordinated execution across up to 300 parallel sub-agents with 4,000 coordinated steps.

Performance Highlights:

• • Humanity’s Last Exam (with tools): 54.0%
  • DeepSearchQA: 92.5% F1 score

Optimal Use Cases: Autonomous research agents, multi-step software refactoring, long-horizon planning with parallel execution, enterprise knowledge management.

DeepSeek V4 Pro (DeepSeek-AI)

Architecture: 1.6T parameter MoE with 49B active parameters, 1M context window, MIT license.

Technical Differentiation: Hybrid thinking/non-thinking architecture dynamically selects between explicit chain-of-thought reasoning and direct response generation based on task complexity, optimizing for both accuracy and latency.

Performance Highlights:

• • GDPval-AA: 1554 ELO (leads open-weight models on agentic real-world tasks)
  • GPQA Diamond: 90.5%, MMLU-Pro: 87.5%

Optimal Use Cases: Complex knowledge work, legal/medical document analysis with extended context, financial modeling with 1M token context, scientific research requiring multi-step reasoning.

GLM 5.1 (Z.ai)

Architecture: 744B parameter MoE with 40B active parameters, 200K context window, MIT license.

Technical Differentiation: Engineered for long-horizon autonomous tasks, maintaining effectiveness over hundreds of reasoning rounds and thousands of tool calls. Incorporates Multi-head Latent Attention (MLA) combined with DeepSeek Sparse Attention (DSA), reducing deployment costs while preserving long-context capacity.

Performance Highlights:

• • Terminal-Bench 2.0: 63.5% (significant improvement over GLM-5 at 56.2)
  • AIME 2026: 95.3%, Autonomous Execution: 8-hour continuous operation

Optimal Use Cases: Long-running software engineering projects, complex system optimization, autonomous DevOps workflows, multi-stage data pipeline optimization.

MiniMax M2.7 (MiniMax AI)

Architecture: 230B parameter MoE with 10B active parameters, 200K context window, non-commercial license.

Technical Differentiation: Demonstrates reduced hallucinations and enhanced real-world task execution. Maintains 97% skill adherence rate across 40+ complex skills exceeding 2,000 tokens each.

Performance Highlights:

• • GDPval-AA: 1495 ELO (highest among open-source models)
  • VIBE-Pro: 55.6% (end-to-end project delivery)

Optimal Use Cases: End-to-end project delivery across platforms, professional office automation, enterprise productivity workflows, complex environment interaction.

DeepSeek V4 Flash (DeepSeek-AI)

Architecture: 284B parameter MoE with 13B active parameters, 1M context window, MIT license.

Technical Differentiation: Efficiency-optimized variant for high-throughput scenarios where latency matters more than absolute accuracy. 5.6x smaller than V4 Pro but maintains competitive performance across most benchmarks.

Performance Highlights:

• • Artificial Analysis Intelligence Index: 47
  • LiveCodeBench Pass@1: 91.6%

Optimal Use Cases: High-volume customer service, real-time code completion, batch document processing, interactive low-latency applications.

These additions demonstrate Dell Technologies’ commitment to supporting open-source models, giving enterprises choice based on specific use cases.

Dell Enterprise Hub: Where open-source meets enterprise reality

Dell Enterprise Hub represents the convergence of open-source innovation and enterprise-grade infrastructure through a comprehensive approach:

Multi-Platform Optimization

Ready-to-deploy on Dell PowerEdge servers powered by the latest accelerators from both NVIDIA (B300, H200, H100, L40S, RTX PRO 6000) and AMD (Instinct MI300X, MI355X) — giving enterprises the flexibility to choose the right silicon for their workload, budget and software ecosystem.

Enterprise-First Security Architecture

Security is built into every layer: all models are scanned for malware and unsafe serialization, Docker images are regularly audited via AWS Inspector, container integrity is verified through SHA384-signed checksums and access is governed through standardized Hugging Face tokens.

Decoupled Architecture for Lifecycle Management

Dell Enterprise Hub’s container versioning system decouples inference containers from model weights, allowing enterprises to pin stable container tags in production, pull or pre-cache weights dynamically (including for air-gapped environments) and update inference engines independently — delivering the version control, flexibility and maintainability that production AI pipelines require.

Goodput Scenarios

Traditional LLM benchmarks focus on raw throughput (requests/tokens per second), but real-world enterprise applications demand more — they must meet specific service level objectives (SLOs) around context length and concurrency. To address this, Dell Enterprise Hub measures performance using Goodput: the actual throughput a system delivers while meeting those SLOs. Each model is optimized across three deployment scenarios — Balanced (a versatile middle ground between context and concurrency), High Concurrency (maximizes requests per second for short-context, high-volume workloads) and Long Context (prioritizes extensive input handling at the expense of concurrency)- giving enterprises a practical, workload-aligned starting point rather than a one-size-fits-all configuration.

The Dell AI SDK: From days to minutes

The SDK automatically matches models to Dell hardware, generates optimal deployment configurations, handles GPU memory allocation and applies platform-specific optimizations — no Docker or Kubernetes expertise required.

Conclusion

The launch of Kimi K2.6, DeepSeek V4 Pro, GLM 5.1, MiniMax M2.7 and DeepSeek V4 Flash — alongside recently added Kimi K2.5, Trinity Large, Cohere Transcribe, Gemma-4 and Mistral-Small-4 — represents a fundamental new capability for enterprise. Organizations can now deploy superior capabilities, with complete control over data, architecture and deployment infrastructure, particularly when combined with PowerEdge XE9780’s purpose-built AI infrastructure and Dell Enterprise Hub’s optimized deployment experience.

^*Learn how Dell AI Factory with NVIDIA delivers a comprehensive and secure AI solution customizable for any business.