NVIDIA, partners release wave of open AI models for local inference

Throughout August, NVIDIA is hosting what it calls a special-edition Local AI celebration, releasing open-source models, software, and developer tools optimized for local execution on consumer and enterprise hardware.

On August 11, the company and partners unveiled multiple models. Meta released Muse Glimmer, a 30-billion-parameter dense open-weight model with a 120,000+ token context window, purpose-built for coding and local agentic AI. On an NVIDIA RTX 5090, it delivers over 200 tokens per second, enabling what NVIDIA describes as "always-on agents" to process data locally and handle complex multistep tasks on a single system. Developers can build custom agents, run private data processing, handle credentials and API keys locally, complete multistep tasks, and sustain long-running workflows without cloud dependency. The model runs on RTX PCs, DGX Spark, DGX Station, and NVIDIA Jetson, and can be fine-tuned locally using NVIDIA NeMo Automodel.

DeepSeek released DeepSeek-V4-Flash, a 284-billion-parameter mixture-of-experts model with 13 billion active parameters and a 1 million-token context window, runnable locally on NVIDIA DGX Station using community-built GGUF versions. Thinking Machines Lab's Inkling-Small is a 276-billion-parameter multimodal model with native reasoning across text, images, and audio; it activates just 12 billion parameters per token and runs on a single DGX Station or two DGX Spark systems. Poolside AI launched Laguna S 2.1, a 118-billion-parameter agentic coding model that works through hours-long tasks, with NVFP4 checkpoints enabling deployment on a single DGX Spark without sacrificing accuracy.

For video and media, LTX released LTX-2.5, a state-of-the-art open-world video generation model featuring multishot support (generating sequences spanning multiple cuts while maintaining continuity), an upgraded diffusion video decoder, and a new prompt enhancer. On an NVIDIA RTX 6000 PRO GPU, LTX-2.5 delivers up to 20% faster performance and 40% memory savings. Alibaba released Wan-Animate-2, a 14-billion-parameter open-weight model for motion transfer from a driving video to a static character; it generates up to 16× faster on NVIDIA RTX PRO 5000 Blackwell and 26× faster on RTX 5090 compared to Apple M3 Ultra. MiniMax-H3, a 33-billion-parameter open-weights video generation model that produces natively synchronized stereo audio from text, images, video, or audio, is available through ComfyUI.

NVIDIA expanded its Nemotron 3 family with Nemotron 3.5 Lightning, a customizable open 30-billion-parameter mixture-of-experts model for always-on agents. It delivers up to 4× faster token generation and 30% faster time to completion compared to open models in its class. Developers can fine-tune it on private examples to write in a preferred style, learn specialized knowledge (photography, gaming, 3D design), or follow coding conventions. NVIDIA collaborated with vLLM, Ollama, llama.cpp, and LM Studio to support NVFP4 and GGUF formats, with Unsloth providing day-one optimized models. Nemotron 3.5 Lightning runs on NVIDIA RTX PCs, DGX Spark, OEM GB10 systems, NVIDIA Jetson, and scales to RTX PRO workstations, DGX Station, and GB300 systems.

On the software side, Unsloth launched Unsloth Desktop on Monday. The fully open-source app combines local model inference, image and video diffusion, fine-tuning, agent integrations, web research, and code execution in a single application, positioned as the first desktop app to both train and run AI models locally. NVIDIA rolled out updates to its Sync app, including a Cluster Assistant that automates configuration of two or more DGX Spark systems as a high-speed cluster via NVIDIA ConnectX-7 ports; Sync detects connected systems, provides secure remote access through Tailscale, and routes workloads across nodes. Arriving later in August are Google Chrome as a native ARM64 Linux build on DGX Spark (installed in a single click) and a new NVIDIA Sync Resource Monitor offering real-time and historical CPU and GPU usage views across individual systems or entire clusters, with marquee zooming for troubleshooting.

NVIDIA also released Cosmos 3 Edge, a 4-billion-parameter open world model for robotics, autonomous vehicles, and vision AI, running on device on DGX Spark and NVIDIA Jetson at a quarter the size of Cosmos 3 Nano. Throughout the updates, the company emphasizes that these models and tools target developers and AI enthusiasts seeking to build, customize, and run capable agents locally without cloud dependency.