MSI has announced that its EdgeXpert desktop AI supercomputer, built on the NVIDIA DGX Spark platform, will launch a new 64GB memory configuration, scheduled to go on sale on October 23, 2026. The new version features the same NVIDIA GB10 Grace Blackwell superchip as the 128GB version, delivers 1 petaFLOP FP4 AI compute, and can run mainstream open-source models with up to 100 billion parameters within enterprise firewalls. The 64GB version and the existing 128GB version form a dual-configuration lineup; the former targets large-scale deployment across departments and multiple sites, while the latter remains at headquarters AI centers for R&D and fine-tuning.
From chatbots to round-the-clock agents
Enterprise AI is evolving from single-turn Q&A chatbots into AI agents that can autonomously execute tasks. Agents must run 24/7, handle multi-step reasoning, and directly access on-premises data, creating infrastructure requirements completely different from the past. EdgeXpert 64GB is positioned as an edge deployment node for around-the-clock Agentic AI: executing close to where data is generated, keeping sensitive data behind the firewall, and balancing data sovereignty with compliance requirements.
EdgeXpert 64GB’s core applications span three industries: in smart manufacturing, a visual quality inspection agent detects product defects in real time and triggers corrective actions, integrating an enterprise knowledge base and a vision-language model. It performs inspection with standard USB or industrial cameras in open factory environments, and operator feedback enables on-site model fine-tuning. The solution won the CES 2026 Enterprise Technology Innovation Award and the Embedded World 2026 Embedded Vision Award. In finance and legal scenarios, fraud investigation and compliance audit agents analyze transaction details and generate reports, corresponding to AIPLUX Legal AI Suite’s contract review, structured extraction, and legal compliance knowledge reasoning, all executed on-premises. In healthcare and retail, clinical record summarization and store inventory scheduling agents run in place on the data, with zero leakage of sensitive data. For voice scenarios, there is also a fully offline speech translation solution in collaboration with Ubestream Taiwan, covering the complete process of speech recognition, translation, and speech synthesis.
A two-tier path of R&D at headquarters and deployment at local sites.
Huang Wen-hui, Vice President of MSI’s Industrial Computer Business Unit, said, “Agentic AI has moved from proof of concept to actual production, fundamentally changing enterprises’ infrastructure needs. The 128GB EdgeXpert can continue to meet R&D teams’ needs to run 200-billion-parameter models or fine-tune 70-billion-parameter models, while the all-new 64GB version is designed specifically for deployment-first enterprises, helping them significantly optimize deployment costs as they scale agents across departments and multiple sites.” A typical deployment path is for the headquarters AI center to first use the 128GB version for prompt engineering and initial validation, then roll out the 64GB version at scale to branch offices, factory floors, retail stores, and highly regulated environments such as hospitals and financial institutions.
The technical premise for the dual-configuration to hold is that the platform is exactly the same. The official blog emphasizes that the 64GB version retains the same GB10 superchip and the same platform; only the memory capacity changes. The scale-up path has also been validated: two 128GB systems interconnected via ConnectX can handle models with up to 405 billion parameters, while the 64GB version also supports multi-node stacking, so a small cluster at a single site can grow as needed. After a team completes prompt engineering and evaluation on a 128GB model, the same containerized agent software can be pushed directly to 64GB edge nodes without rewriting any code, eliminating software fragmentation caused by different architectures between development machines and production machines. At the same time, the architecture natively supports multi-node scaling: a single unit can be deployed independently at a remote site, and performance scales linearly when multiple units are interconnected. NVIDIA Sync helps teams build clusters made up of multiple interconnected systems and monitor CPU and GPU operations through a single desktop application.
What models can actually fit in 64GB?
The official blog directly answers the most practical question. After deducting about 8GB for system overhead, the 64GB version has about 56GB left for model weights and KV cache. Memory capacity directly determines the context length and KV cache space for multi-turn inference. Even when loading dense models above 30B or MoE models, the system still reserves 35GB to 41GB of pure KV cache, ensuring token generation speed and contextual memory depth when agents run continuously over long periods. Paired with the GB10 chip’s 1 PFLOP FP4 compute, the official conclusion is that the combination of performance, thermal efficiency, and unit cost exactly matches the real-world needs of mass-production scenarios such as factory quality inspection, financial document processing, and clinical voice agents.
Product Summary and Launch Information
On the hardware side, the EdgeXpert 64GB features the NVIDIA GB10 Grace Blackwell superchip, which includes a Blackwell GPU and a 20-core Arm CPU, with NVLink-C2C interconnecting CPU and GPU memory. The 64GB LPDDR5x unified memory supports models with up to 100 billion parameters. The cooling is designed for always-on agents, delivering sustained high-performance output. The chassis maintains a 1.2-liter, 1.2-kg desktop design, focusing on ultra-quiet cooling. Networking includes a 10GbE port and a 200GbE NVIDIA ConnectX-7 network card, supporting system cluster deployment. NVLink-C2C bandwidth is 5 times that of PCIe 5.0, ensuring ultra-fast data access and transfer. Storage offers 1TB or 4TB NVMe self-encrypting options, plus Wi-Fi 7, Bluetooth 5.3, HDMI 2.1a, and four USB 3.2 Type-C ports. The chassis dimensions are 151 × 151 × 52 mm. Software comes preloaded with NVIDIA DGX OS and the complete NVIDIA AI software stack, ready out of the box, with fully on-premises inference enabling zero data leakage, full compliance, and optimized deployment costs.
The EdgeXpert 64GB version will launch on October 23, 2026. MSI has previously accumulated deployment cases including confidential legal and intellectual property management, smart factory operations, offline speech translation, and autonomous AI agent workflow reasoning. With this new version, enterprises essentially gain a complete path from headquarters R&D to mass production at branch sites: the large ones are reserved for training and fine-tuning, while the small ones are rolled out to every industrial site that needs agents. For more specifications and enterprise deployment information, refer to the MSI EdgeXpert product page.
Source: KOCPC Chinese