AI อะไรเนี่ย

Model

NVIDIA Unveils Nemotron 3 & Open Models for Local AI Agents on RTX PCs

NVIDIA Unveils Nemotron 3 & Open Models for Local AI Agents on RTX PCs

Hey there, AI enthusiasts! NVIDIA just dropped some exciting news at GTC that's set to revolutionize how we interact with artificial intelligence. The big takeaway? Powerful, private AI agents are coming to your desktop, running locally on NVIDIA RTX PCs and DGX Spark supercomputers. This isn't just about faster processing; it's about giving you unprecedented control, privacy, and performance for your AI workflows.

The spotlight shines brightly on new open models from the NVIDIA Nemotron 3 family, alongside significant optimizations for popular models like Mistral and Qwen. Plus, NVIDIA introduced NemoClaw, an open-source stack designed to enhance the OpenClaw experience on NVIDIA hardware. It’s a game-changer for anyone looking to harness the full potential of local AI. You can dive deeper into all the announcements in the NVIDIA Blog Article.

What It's All About: Next-Gen Local AI

NVIDIA's latest push is all about bringing cloud-level AI capabilities right to your personal devices, transforming them into "agent computers." This means your AI can access richer user context, interact with local tools, and automate tasks while keeping your data private.

Leading the charge are the new additions to the NVIDIA Nemotron 3 family of open models:

  • Nemotron 3 Nano 4B: This compact yet capable model is perfect for RTX AI PCs and devices with limited resources. Imagine building action-taking conversational agents for games or apps that run seamlessly on your hardware, demanding minimal VRAM.
  • Nemotron 3 Super 120B: Designed for the most complex agentic AI systems, this 120-billion-parameter open model (with 12 billion active parameters) is optimal for DGX Spark or NVIDIA RTX PRO workstations. It scored an impressive 85.6% on PinchBench, making it a top contender for OpenClaw performance in its class.

Beyond Nemotron 3, NVIDIA also announced crucial optimizations for other leading open models:

  • Mistral Small 4: A robust 119-billion-parameter model tailored for general chat, coding, and various agentic tasks, offering a unified and ultra-efficient solution.
  • Alibaba's Qwen 3.5 Models (27B, 9B, 4B): These models now boast NVIDIA optimizations, supporting vision, multi-token prediction, and an astonishing 262,000-token context window. The dense 27-billion-parameter version, for instance, truly shines when paired with an RTX 5090 GPU.

To facilitate secure and optimized local agent deployment, NVIDIA unveiled NemoClaw, an open-source stack for OpenClaw. This stack helps deploy optimizations for OpenClaw on NVIDIA devices, featuring Nemotron local models for private inference and OpenShell runtime for safer execution of "claws."

Why Local AI Agents are a Game-Changer

The shift towards local AI agents addresses several key concerns for users and developers:

  • Privacy and Cost Savings: By running models locally, your data stays on your device, enhancing privacy. Furthermore, it eliminates token costs associated with cloud-based inference, making AI more accessible and economical for sustained use.
  • Unrivaled Performance: Devices like the DGX Spark desktop AI supercomputer come with a massive 128GB of unified memory, capable of supporting models with over 120 billion parameters. NVIDIA RTX GPUs further accelerate inference, ensuring a smooth and responsive AI experience.
  • Democratizing Development: With tools like Unsloth Studio, fine-tuning models for specific data and use cases has never been easier. This web-based interface simplifies the complex fine-tuning process, allowing more developers to customize and improve model accuracy for their agentic workflows.

Moreover, creative AI workflows are getting a significant boost with RTX-optimized models:

  • Lightricks' LTX 2.3: This state-of-the-art audio-video model now supports NVFP4 and FP8 distilled models, accelerating performance by an impressive 2.1x. You can learn more about the LTX 2.3 release here.
  • Black Forest Lab's FLUX.2 Klein 9B: An update to this image editing model, including an FP8 version optimized for RTX GPUs, speeds up image editing by up to 2x. Check out the FLUX.2 Klein 9B model details.

Getting Started: Where You Get These Tools

Ready to experiment with these groundbreaking local AI agents?

  • Model Availability: The Nemotron 3 family, Qwen 3.5, and other optimized models are readily available through popular platforms like Ollama, LM Studio, and llama.cpp.
  • Fine-Tuning: If you're looking to customize models, Unsloth Studio provides an intuitive web-based UI, simplifying the process for over 500 AI models.
  • NemoClaw: Dive into the open-source stack to build more secure and private OpenClaw agents on your NVIDIA-powered devices.

These innovations mark a pivotal moment for personal computing, bringing sophisticated AI capabilities closer to users than ever before. It's an exciting time to be building and interacting with AI!

Read more: NVIDIA's Latest AI Innovations for details and links to NVIDIA's latest AI innovations.