NVIDIA PAIR and Broadcast updates arrive with Llama Windows

Listen to this article

NVIDIA released PAIR 1.0 and Broadcast 2.3 on October 13, alongside the debut of Llama Windows from Hugging Face and the llama.cpp community. The software announcements arrive ahead of RTX Spark laptops, which NVIDIA says will become available on October 16.

The practical changes cover three different jobs. PAIR sends inference requests to available computers on a local network. Broadcast applies AI effects to recordings and live media. Llama Windows gives people a simpler way to download and run models on a Windows PC.

PAIR adds more control over local inference

The PAIR 1.0 release notes add llama.cpp alongside the existing Ollama and LM Studio backends. Users can download compatible GGUF models from Hugging Face and adjust engine launch arguments, including context settings where the engine supports them. NVIDIA also lists validation for RTX Spark hardware.

PAIR routes requests between separate machines. Its documentation explicitly says the cluster does not merge those devices into one virtual GPU. That matters when choosing models because spreading tasks across computers does not by itself remove the memory limits of an individual machine.

There is also a compatibility change for browser clients. PAIR endpoints now follow each inference engine’s allowed origin policy. Some existing browser applications may therefore need their engine configuration updated before they can connect.

Broadcast adds recorded media processing

Broadcast 2.3 introduces a File Processor that applies supported audio and video effects to saved media. Its release notes also describe an MCP server through which local AI agents can change supported settings or process files. Live processing and file processing cannot run at the same time.

The two new video effects remain in beta. Super Resolution supports upscaling to 4K on RTX 3060 hardware and above. Frame Generation requires an RTX 40 Series GPU or newer, with up to 60 frames per second for live video and doubled or quadrupled frame rates for files. These requirements are specific to the effects.

Llama Windows puts local models in the tray

Hugging Face’s separately dated October 13 announcement introduces Llama Windows as a free application for Windows 11. A tray panel helps users choose and download models from the Hub, then load them into llama.cpp. The app also supports Windows on Arm and includes an overlay opened with Alt and Space.

The application can connect to an existing llama.cpp server or start the runtime it manages. Developers can use that local server through an API compatible with OpenAI clients. Model downloads use the shared Hugging Face cache, and the settings allow users to choose another drive.

Some announced products arrive later

NVIDIA’s announcement combines current software with a release schedule. Nemotron 3.5 Bolt 74B is scheduled for October 15, followed by RTX Spark laptops on October 16. The smaller Bolt E4B model and the next RTX Remix update are scheduled for October 23. Those dates should not be read as availability today.

For people already running local AI, the immediate decision is which part of their workflow needs improvement. PAIR addresses distributing work, Broadcast addresses media processing, and Llama Windows addresses model setup. Real performance will still depend on the model, hardware and workload selected.

ByteForward’s reporting on FastFlowLM for AMD NPUs examines another local inference route and its hardware requirements.

Illustrative media workspace photographed by Farhan Srambiyan. CC0 via the WordPress Photo Directory.