Nvidia Expands Open-Weight AI Ecosystem with New Robotics and Multimodal Models
As reported by CNBC, Nvidia has begun rolling out open-weight models following a recent public exchange involving CEO Jensen Huang, marking the company's first open-source release tranche in that window.
The August push—documented in detail on Nvidia's developer blog—spans video, robotics, coding and multimodal families, each parameterized and quantized to map onto the vendor's own workstation and edge silicon tiers.
The model lineup
Cosmos 3 Edge arrives as a 4-billion-parameter open world model for robotics, autonomous vehicles and vision AI, engineered to run on-device on DGX Spark and Jetson platforms. At roughly a quarter the parameter count of Cosmos 3 Nano, it materially shifts the latency and power envelope for edge autonomy stacks where Nano's footprint was previously the binding constraint. Video generation is covered by MiniMax-H3, a 33-billion-parameter open-weights model producing video with natively synchronized stereo audio from text, image, video or audio inputs, distributed through ComfyUI with checkpoints optimized for Nvidia GPUs. Alibaba's Wan-Animate-2 contributes a 14-billion-parameter motion-transfer model that retargets driving video onto static character images, with Nvidia claiming up to 16× generation speedup on RTX PRO 5000 Blackwell (48GB) and 26× on RTX 5090 relative to Apple M3 Ultra.
Agentic and multimodal workloads
Poolside AI's Laguna S 2.1 enters as a 118-billion-parameter open-weight coding model engineered for hours-long task horizons. An NVFP4 quantization checkpoint reduces compute and memory requirements, enabling local execution on a single DGX Spark. DeepSeek refreshed DeepSeek-V4-Flash, a 284-billion-parameter Mixture-of-Experts architecture activating 13 billion parameters per token with a 1-million-token context window, deployable on DGX Station via community-built GGUF versions. The deployment ceiling is set by Thinking Machines Lab's Inkling-Small: 276 billion total parameters, 12 billion activated per token, multimodal across text, images and audio with adjustable reasoning effort. Trained on GB300 NVL72 infrastructure, it executes on a single DGX Station or two paired DGX Spark systems, with an NVFP4 Blackwell-optimized checkpoint available on Hugging Face.
Infrastructure and on-device tooling
The release coincides with a $240 million multi-year agreement between IBM and Together AI to construct a large-scale inference cluster on IBM Cloud, built on Nvidia HGX B300 systems—effectively positioning Blackwell as a hosted inference tier for open-weight models alongside the on-device DGX pathway. Unsloth's desktop release consolidates local inference, image and video diffusion, fine-tuning, agent integration, web research and code execution into a single open-source application, marketed by its developers as the first desktop app to both train and run AI models locally. For practitioners mapping targets, the deployment matrix from the Nvidia rollout is straightforward: DGX Spark for ≤118-billion-parameter NVFP4 inference paths; DGX Station for full-precision GGUF execution of DeepSeek-V4-Flash; Jetson or DGX Spark for Cosmos-class edge workloads; and IBM Cloud's HGX B300 cluster for hosted open-weight inference at scale.