NVIDIA Expands Nemotron and Cosmos Families as Open-Weight Partners Ship Local AI Models

NVIDIA's own open model moves: vision AI and agentic models

According to NVIDIA's Aug. 11 blog post, the company released Cosmos 3 Edge, a 4-billion-parameter open world model built for robotics, autonomous vehicles and vision AI. NVIDIA describes it as a quarter the size of Cosmos 3 Nano, and it runs on device on NVIDIA DGX Spark and NVIDIA Jetson [E1].

The same post states that NVIDIA expanded its Nemotron 3 model family with Nemotron 3.5 Lightning, a customizable open 30-billion-parameter mixture-of-experts (MoE) model designed for always-on agents [E13]. Together, the two releases cover both an edge-deployed vision model and a general-purpose agentic MoE model.

Open-weight agentic models: parameter scale across vendors

NVIDIA's blog post reports that Poolside AI launched Laguna S 2.1, a 118-billion-parameter open-weight agentic coding model built to work through hours-long tasks; an NVFP4 checkpoint lets it run locally on a single DGX Spark with lower compute and memory requirements, which NVIDIA says comes without sacrificing accuracy [E3].

DeepSeek refreshed DeepSeek-V4-Flash, a 284-billion-parameter MoE model with 13 billion active parameters and a 1 million-token context window; per the post, developers can run it locally on an NVIDIA DGX Station using community-built GGUF versions [E4]. Thinking Machines Lab's Inkling-Small is described as a frontier-class, open-weight multimodal model that activates just 12 billion of its 276 billion parameters per token, trained on NVIDIA GB300 NVL72, and runs on a single DGX Station or two DGX Spark systems [E5]. Nemotron 3.5 Lightning's 30-billion-parameter MoE design sits at the smaller end of this group [E13].

Multimodal and video generation: a parameter race

According to the same post, MiniMax released MiniMax-H3, a 33-billion-parameter open-weight model that generates video with natively synchronized stereo audio from text, images, video, audio or a mix of the four [E2]. 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 — human, cartoon, robot or animal — with day-zero support in ComfyUI [E7]. Meta released Muse Glimmer, a 30-billion-parameter dense open-weight model with a context window exceeding 120,000 tokens, which NVIDIA's post says is purpose-built for coding and local agentic AI [E9]. Parameter counts for this group range from MiniMax-H3's 33 billion down to Wan-Animate-2's 14 billion, with Muse Glimmer's 30 billion falling in between.

Local execution and edge deployment

Cosmos 3 Edge's 4-billion-parameter footprint is specifically built to run on device on DGX Spark and Jetson [E1]. NVIDIA's blog post also reports that Unsloth is launching Unsloth Desktop, which the launch materials position as the first desktop app that both trains and runs AI models locally; it combines local model inference, image and video diffusion, fine-tuning, agent integrations, web research and code execution into one open-source desktop application [E6]. Separately, the post states that DGX Spark will get a native ARM64 Linux build of Google Chrome, installable in a single click from the DGX Dashboard [E12] — a convenience feature aimed at making the on-device environment more self-contained.

GPU-accelerated performance: speed, memory, throughput

NVIDIA's post cites several vendor-reported performance comparisons run on its GPUs. Wan-Animate-2 generates up to 16x faster on an NVIDIA RTX PRO 5000 Blackwell (48GB) and up to 26x faster on an NVIDIA RTX 5090 compared with Apple M3 Ultra [E7]. LTX-2.5, optimized for NVIDIA RTX GPUs, DGX Spark and DGX Station systems, delivers up to 20% faster performance and 40% memory savings on an NVIDIA RTX 6000 PRO GPU [E8]. Muse Glimmer delivers over 200 tokens per second on an RTX 5090, which NVIDIA's post says enables always-on agents to process data locally through complex, multistep tasks on a single system [E10]. Nemotron 3.5 Lightning delivers up to 4x faster token generation and 30% faster time to completion compared with open models in its class [E14].

NVIDIA's inference infrastructure: clustering, local environment, routing costs

According to NVIDIA's blog post, the Cluster Assistant in NVIDIA Sync automates configuring two or more DGX Spark systems as a high-speed cluster: developers connect the systems through their NVIDIA ConnectX-7 ports, and NVIDIA Sync configures the network, routes workloads across nodes and monitors system health [E11]. The same post notes DGX Spark's incoming single-click, native ARM64 Chrome install from the DGX Dashboard as part of rounding out the local development environment [E12]. On the cost side, NVIDIA reports that internal benchmarks show NeMo Switchyard — which routes each step of a task across a system of models — helped maintain frontier-level task completion while reducing benchmark completion cost to roughly one-third of using Opus 4.8 alone [E15].

What this means

Across the evidence in NVIDIA's Aug. 11 post, two patterns line up. First, parameter scale varies widely — from Cosmos 3 Edge's 4 billion to DeepSeek-V4-Flash's 284 billion — yet nearly every model in the roundup is paired with a specific NVIDIA local-hardware target: DGX Spark, DGX Station, Jetson or RTX GPUs. Second, the performance figures NVIDIA cites (Wan-Animate-2's up to 26x speedup, LTX-2.5's 40% memory savings, Nemotron 3.5 Lightning's 4x token-generation gain, and NeMo Switchyard's roughly one-third cost reduction versus Opus 4.8 alone) are all benchmarked on NVIDIA's own GPUs and published through NVIDIA's own blog. That means the comparisons describe how third-party open-weight models perform on NVIDIA hardware, as reported by NVIDIA — not independently verified figures.