NVIDIA has unveiled a suite of tools and models designed to make AI agents practical for developers building on local hardware rather than relying solely on cloud services. The centerpiece is NemoClaw, an agent toolkit that works alongside Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model with open weights that can run on consumer GPUs, workstations, and compact systems.

What Makes Local AI Agents Different From Cloud-Based Alternatives?

Local AI agents represent a shift in how developers approach automation tasks. Instead of sending every request to a remote cloud service, local agents handle repeated or specialized tasks directly on a user's device, keeping data private and reducing latency. Nemotron 3.5 Lightning is built for this "always-on" use case, allowing developers to fine-tune the model for specific applications without uploading sensitive information to external servers.

The model runs across a wide range of NVIDIA hardware, from consumer RTX graphics cards to DGX Spark systems and Jetson devices. Major hardware partners including Acer, ASUS, Dell, HP, Lenovo, and others are supporting Blackwell-based systems that can run the software, making the toolkit accessible to developers with varying budgets and infrastructure.

How to Deploy AI Agents on Your Local Hardware

• Choose Your Hardware: Nemotron 3.5 Lightning runs on RTX consumer GPUs, DGX Spark compact systems, Jetson edge devices, and workstations, so you can start with existing equipment or invest in a single high-end GPU.
• Use NemoClaw for Agent Building: The toolkit provides the software framework needed to construct agents that handle specific workflows, with support from popular deployment tools like vLLM, Ollama, llama.cpp, and LM Studio.
• Fine-Tune for Your Use Case: NVIDIA's NeMo Automodel allows developers to customize the model for specialized tasks while keeping training data on the device, avoiding cloud uploads.
• Optimize Costs With NeMo Switchyard: This open-source routing library directs different stages of a task to different models based on speed, accuracy, and cost, reducing token expenses compared to using a single top-tier model for everything.

NVIDIA introduced NeMo Switchyard, an open-source routing library that addresses a growing concern among businesses deploying generative AI: the cost of sending every task to a premium model. Internal benchmarks showed that routing workloads across multiple models with Switchyard reduced benchmark completion cost to roughly one-third of using a top-tier model alone while maintaining frontier-level task completion.

What Other Models and Tools Are Part of This Release?

The announcement extends beyond NemoClaw and Nemotron 3.5 Lightning. NVIDIA highlighted support for several other open-weight models designed for local deployment, including Cosmos 3 Edge, a 4-billion-parameter world model for robotics and vision applications, and MiniMax-H3, a 33-billion-parameter model that generates video and stereo audio.

Meta's Muse Glimmer, a 30-billion-parameter open-weight model designed for coding and local agent tasks, can run on a single GeForce RTX 5090 while handling long context windows and multistep workflows. NVIDIA also updated support for DeepSeek-V4-Flash and Thinking Machines Lab's Inkling-Small, both large models that can run with reduced active parameter counts through optimized checkpoints.

A new desktop application called Unsloth Desktop is launching as a fully open-source tool for local model inference, training, diffusion workloads, agent integrations, web research, and code execution. The software combines local AI training and inference in one application, eliminating the need to juggle multiple tools.

How Does NVIDIA's Sync Software Enable Clustering?

For developers who need more computing power than a single system provides, NVIDIA updated its NVIDIA Sync software to let users cluster multiple DGX Spark systems together. The application automatically detects connected systems, provides remote access, and launches applications across one or more DGX Spark units. A Cluster Assistant within the software configures two or more systems as a high-speed cluster through ConnectX-7 ports, allowing developers to run larger models like GLM 5.2 and DeepSeek V4 Flash when a single system does not provide enough memory or throughput.

NVIDIA also disclosed two additional DGX Spark software updates: a native ARM64 Linux build of Google Chrome and a Sync Resource Monitor for system-level usage tracking, giving developers better visibility into how their hardware is being used.

Why Does This Matter for Developers and Enterprises?

The broader message from NVIDIA is that larger open models can increasingly run on desktop or compact workstation hardware without requiring expensive cloud infrastructure. This democratizes access to advanced AI capabilities and addresses privacy concerns by keeping data local. For enterprises, the cost savings from routing workflows across multiple models rather than always using premium services could be substantial, especially for organizations running thousands of agent tasks daily.

The toolkit also reflects a growing trend toward open-source and open-weight models in the AI industry. By supporting popular deployment frameworks like vLLM, Ollama, and llama.cpp, NVIDIA is making it easier for developers to experiment with different models and avoid vendor lock-in. The effort now spans coding agents, robotics, media generation, and multimodal systems across hardware ranging from consumer GPUs to clustered desktop appliances.