NVIDIA released Nemotron 3.5 Lightning, an open 30-billion-parameter mixture-of-experts model designed to run local agent workflows on personal computers, workstations, and edge hardware. NVIDIA reported that the model delivers up to four times faster token generation and 30 percent faster task completion than comparable open models in its class.
Developers can deploy Nemotron 3.5 Lightning locally on RTX PCs, DGX Spark systems, OEM GB10 units, and Jetson edge devices, as well as RTX PRO workstations and DGX Station hardware. The company collaborated with vLLM, Ollama, llama.cpp, LM Studio, and Unsloth to provide NVFP4 and GGUF format checkpoints. Alongside the model, NVIDIA published NeMo Switchyard on GitHub, an open-source routing tool that directs steps in an agent workflow to specific models based on speed, accuracy, and cost.
Agent and Coding Models
Meta released Muse Glimmer, a 30-billion-parameter dense open weight model with a context window exceeding 120,000 tokens. The model runs locally on GeForce RTX PCs, DGX Spark, and Jetson hardware, reaching over 200 tokens per second on a single RTX 5090 GPU.
NVIDIA also added day-zero support for Qwen3.8-27B, a 27-billion-parameter local coding model designed as a companion to Qwen3.8-Max. Using multi-token prediction, Qwen3.8-27B reaches 131 tokens per second on a single RTX 5090. Poolside AI launched Laguna S 2.1, a 118-billion-parameter coding model with an NVFP4 checkpoint that runs on a single DGX Spark, while DeepSeek refreshed DeepSeek-V4-Flash with a 1-million-token context window across 284 billion total and 13 billion active parameters.
Media Generation and System Tools
LTX introduced LTX-2.5, an open video generation model with multishot sequencing and a prompt enhancer powered by Gemma4 text encoders. The model delivers up to 20 percent faster performance and 40 percent memory savings on NVIDIA RTX 6000 PRO hardware. Alibaba released Wan-Animate-2, a 14-billion-parameter animation model that generates up to 22 times faster on an RTX 5090 than on Apple's M3 Ultra.
Unsloth launched Unsloth Desktop, an open-source application that trains and runs models locally. NVIDIA also updated its Sync app for macOS and Windows, adding a Cluster Assistant that connects multiple DGX Spark units across ConnectX-7 ports without manual network configuration. Later in August, DGX Spark will receive a native ARM64 build of Google Chrome alongside a built-in Sync Resource Monitor for cluster usage tracking.
