AI agents are shedding their cloud chains. The latest wave of models and platforms now run entirely on local hardware. This means faster responses, no data leaks, and real control over your machine.

H company rolled out Holo3.1, an upgrade focused on local computer-use agents that run on desktops and mobile devices. It supports multiple environments—web, desktop, and Android—with native function-calling and quantized checkpoints for lightweight, private inference. Smaller models start at 0.8 billion parameters, scaling to a 35-billion-parameter version for power users.

Holo3.1’s quantized formats—FP8, Q4 GGUF, NVFP4—slash memory use and boost speed. On NVIDIA hardware, these optimizations double throughput and halve step times. This lets agents automate workflows directly on consumer PCs without cloud dependencies.

Meanwhile, Microsoft’s Foundry Local hit general availability. This runtime embeds AI inference inside your app with no background servers or daemons. It detects hardware accelerators automatically—from GPUs to NPUs—and supports SDKs in Python, JavaScript, C#, and Rust. Foundry Local also offers an OpenAI API-compatible interface, making it trivial to switch between local and cloud models.

Foundry Local’s model catalog focuses on small-to-medium open models optimized for edge deployment. It includes Microsoft’s Phi-4 series, Qwen 2.5 variants, Mistral 7B, and Whisper for audio transcription. The platform targets regulated industries where data sovereignty and zero egress matter most.

NVIDIA is pushing local AI agents through new RTX Spark PCs and DGX systems. RTX Spark PCs pack up to 1 petaflop of AI compute and 128GB unified memory, designed to run personal AI agents on-device. NVIDIA also introduced DGX Station for Windows to bring high-end inference to desktops.

The company’s software stack integrates security controls via OpenShell and Microsoft’s new Windows security primitives. This safeguards local agent operations, enabling fine-grained access control. Early adopters like Hermes Agent and OpenClaw are building Windows applications that harness these features.

NVIDIA’s ecosystem update spans creative tools too. Adobe is rearchitecting Photoshop and Premiere for RTX Spark, enabling GPU-accelerated compositing, live filters, and AI-assisted editing. Blender gains DLSS 4.5 for real-time path-tracing viewports. RTX Video Frame Generation aims to double or quadruple video frame rates in real time.

Behind the scenes, NVIDIA optimized popular open-source projects like llama.cpp and vLLM. These efforts doubled inference speeds on Qwen 3.5 and 3.6 models on DGX Spark. Multi-GPU tensor parallelism now boosts memory and compute efficiency, benefiting local generative AI workflows.

On the model front, NVIDIA and HuggingFace unveiled Cosmos 3, a unified omni-model for physical AI. It merges perception, reasoning, and action into one transformer architecture. Cosmos 3 understands instructions, processes live video feeds, and outputs robot control commands in a single pass. It comes in 8-billion and 32-billion parameter versions, suitable for workstation and data center hardware respectively.

Cosmos 3’s multimodal design replaces clunky robotics pipelines that split vision, language, and control into separate models. It scored 92% success on MetaWorld assembly tasks and 78% on complex manipulation benchmarks, outpacing prior open-source systems by double digits.

The open-source release includes model weights, training code, and synthetic datasets to fine-tune on custom robotics tasks. Developers can adapt Cosmos 3 to various platforms — from industrial arms to humanoid robots — with as few as 100 examples.

The AI landscape is shifting. Cloud-only AI is no longer the default. Privacy-conscious enterprises and developers want local agents that run fast and securely on their own hardware. Hardware vendors and software makers are answering the call with optimized models, runtimes, and developer tools.

This year’s breakthroughs point to a future where your PC or mobile device handles complex AI workflows independently. That means less waiting, more privacy, and AI agents that actually live where you do.