NVIDIA is highlighting the expanding open source ecosystem for local AI agents, featuring multiple new models and tools

Throughout August, NVIDIA is celebrating partners and open source communities advancing local AI, including new models, applications and tools. The company is promoting its RTX GPUs alongside the software ecosystem, libraries and educational resources supporting local AI development.

The Qwen3.8-27B open model serves as the local companion to Qwen3.8-Max. As a 27-billion-parameter model sized for a single GPU, it is built for responsive coding workflows with local files, tools and project context. Qwen3.8-27B reaches 131 tokens per second on a single GeForce RTX 5090 with multi-token prediction, enabling developers to keep sensitive code and proprietary data on their own systems. Day-zero support is available on NVIDIA RTX PRO GPUs and the GeForce RTX 5090. Developers can use llama.cpp, Ollama, Unsloth and LM Studio Bionic, all supporting Qwen3.8-27B on NVIDIA RTX GPUs.

Beyond personal AI systems, NVIDIA Jetson brings day-zero support for Qwen3.8-27B to the edge. NVIDIA recently released Cosmos 3 Edge, a 4-billion-parameter open world model for robotics, autonomous vehicles and vision AI. At a quarter the size of Cosmos 3 Nano, it runs on NVIDIA DGX Spark and NVIDIA Jetson.

MiniMax-H3 is a 33-billion-parameter open weights model that generates video with natively synchronized stereo audio from text, images, video, audio or combinations thereof. Creators can access it through ComfyUI and run it locally with checkpoints optimized for NVIDIA GPUs.

Poolside AI launched Laguna S 2.1, a 118-billion-parameter open weight agentic coding model for hours-long tasks. An NVFP4 checkpoint enables developers to run the model locally on a single NVIDIA DGX Spark with lower compute and memory requirements.

DeepSeek recently refreshed DeepSeek-V4-Flash, a 284-billion-parameter mixture-of-experts model with 13 billion active parameters and a 1 million-token context window. Developers can run it locally on an NVIDIA DGX Station using community-built GGUF versions.

Thinking Machines Lab's Inkling-Small is a frontier-class open weight multimodal model with native reasoning across text, images and audio. The 276-billion-parameter model activates just 12 billion parameters per token and runs on a single DGX Station or two DGX Spark systems. An NVFP4 checkpoint optimized for NVIDIA Blackwell is available on Hugging Face.

Unsloth launched Unsloth Desktop, bringing local model inference, image and video diffusion, fine-tuning, agent integrations, web research and code execution into one fully open source desktop application. It is the first desktop app that both trains and runs AI models locally.

Alibaba released Wan-Animate-2, a 14-billion-parameter open weight model that transfers motion and facial expressions from a driving video onto a static character image. With day-zero support in ComfyUI, it generates up to 26 times faster on NVIDIA RTX PRO 6000 Blackwell and 22 times faster on an NVIDIA RTX 5090 compared to Apple M3 Ultra.

LTX-2.5 is the latest open-world video generation model from LTX's research team. New multishot support enables creators to generate sequences spanning multiple cuts while maintaining continuity. A new diffusion video decoder improves visual fidelity and reduces artifacts. The prompt enhancer, featuring Gemma4 E2B and a custom Gemma4 12B text encoder, provides precise prompt adherence. LTX-2.5 is optimized for NVIDIA RTX GPUs, DGX Spark and DGX Station systems. On an NVIDIA RTX 6000 PRO GPU, LTX-2.5 delivers up to 20 percent faster performance and 40 percent memory savings.

Meta released Muse Glimmer, a 30-billion-parameter, dense, open weight model with a 120-thousand-plus token context window, purpose-built for coding and local agentic AI. Optimized for NVIDIA GeForce RTX PCs, NVIDIA DGX Spark, DGX Station and NVIDIA Jetson, Muse Glimmer delivers over 200 tokens per second on RTX 5090, enabling always-on agents to process data locally and work through complex multistep tasks. Its dense architecture and hybrid attention keep processing and memory demands manageable as agents take on longer tasks, use tools and maintain context across multiple steps. Developers can run Muse Glimmer with vLLM for text generation, image, reasoning and tool use, or llama.cpp for text and image workloads using BF16 and quantized GGUF checkpoints.