NVIDIA highlights a major expansion of the local AI ecosystem this August, showcasing new open-source models, developer

Throughout August, NVIDIA is celebrating the open source ecosystem advancing local artificial intelligence, highlighting emerging models, applications, tools, accelerated computing libraries, and educational resources. This special-edition Local AI blog series will feature weekly updates on these developments. Users can follow NVIDIA RTX Spark on X, Instagram, TikTok, and Facebook, subscribe to the NVIDIA Local AI newsletter, and follow NVIDIA Workstation on LinkedIn and X.

Perplexity introduced Portable Computer, a local agent app for the NVIDIA DGX Spark that enables private workflows without counting against token limits. Designed for productivity workers, it integrates with Google Drive, Gmail, Slack, and GitHub, and allows switching between cloud models for complex tasks and local models for daily use. The app leverages a specially post-trained Qwen 3.8 27B model and offers one-click local inference setup optimized for NVIDIA hardware. Support for GeForce RTX and RTX PRO GPUs, Windows, and DGX Station are in development, alongside a forthcoming fine-tuned Nemotron 3.5 Lightning variant for ultra-fast responses.

NVIDIA highlighted day-zero support for the 27-billion-parameter Qwen3.8-27B open model on RTX GPUs, including RTX PRO and GeForce RTX 5090. Sized for a single GPU, it serves as a local companion to Qwen3.8-Max, enabling developers to keep sensitive code and project context on-premises. Multi-token prediction optimization delivers 131 tokens per second on a single GeForce RTX 5090. The model is compatible with llama.cpp, Ollama, Unsloth, and LM Studio Bionic, and receives additional optimizations for DGX Spark and DGX Station, including upcoming NVFP4 support. Edge deployment is also enabled on NVIDIA Jetson devices.

Several other open-weight models were announced for local deployment. Cosmos 3 Edge is a 4-billion-parameter open world model for robotics and vision AI, running on DGX Spark and Jetson. MiniMax-H3 is a 33-billion-parameter model generating video and synchronized stereo audio from mixed media inputs, accessible via ComfyUI with GPU-optimized checkpoints. Poolside AI launched Laguna S 2.1, an 118-billion-parameter agentic coding model that handles long-running tasks, with an NVFP4 checkpoint allowing operation on a single DGX Spark. DeepSeek refreshed DeepSeek-V4-Flash, a 284-billion-parameter MoE model with 13 billion active parameters and a 1 million-token context window, runnable on DGX Station via community GGUF versions. Thinking Machines Lab released Inkling-Small, a 276-billion-parameter multimodal model activating 12 billion parameters per token, trained on GB300 NVL72 and deployable on a single DGX Station or two DGX Spark systems, with an NVFP4 Blackwell checkpoint available on Hugging Face.

New software and tooling continue to streamline local AI workflows. Unsloth Desktop, launched recently, is a fully open source application combining local inference, image and video diffusion, fine-tuning, agent integrations, web research, and code execution, marking the first desktop app capable of both training and running AI models locally. Alibaba released Wan-Animate-2, a 14-billion-parameter model that transfers motion and facial expressions from driving videos to static images, achieving up to 26x faster generation on NVIDIA RTX PRO 6000 Blackwell and 22x faster on an NVIDIA RTX 5090 compared to Apple M3 Ultra, with day-zero support in ComfyUI.

LTX-2.5 represents the latest state-of-the-art open-world video generation model, offering improved quality, prompt adherence, and character consistency. It features multishot support for continuous multi-cut sequences, an upgraded diffusion decoder to reduce artifacts, generative editing capabilities, and a prompt enhancer powered by Gemma4 E2B and a custom Gemma4 12B text encoder. Optimized for NVIDIA RTX GPUs, DGX Spark, and DGX Station, it achieves up to 20% faster performance and 40% memory savings on an NVIDIA RTX 6000 PRO using NVFP4, FastVideo, and ready-to-use ComfyUI workflows. Finally, Meta released Muse Glimmer, a 30-billion-parameter dense open weight model with a 120K+ context window purpose-built for coding and local agentic AI. Optimized for GeForce RTX PCs, DGX Spark, DGX Station, and NVIDIA Jetson, it delivers over 200 tokens per second.