Four releases from NVIDIA, Mistral and Liquid AI target robot control, driving-data workflows, content moderation and on-device agents, showing how open weights now serve sharply different jobs.
Artificial Intelligence··Evening
From robot arms to driving data
NVIDIA's two releases open separate working paths for physical AI. The Cosmos 3 family comes as Edge with 4 billion parameters, Nano with 16 billion and Super with 64 billion. Edge is intended for on-device deployment and Nano for workstation serving; two policies post-trained from the Nano and Edge checkpoints target Franka Panda arms with Robotiq grippers. Given language instructions and multi-camera observations, the policies generate robot-action trajectories. NVIDIA distributes the weights alongside datasets, the post-training recipe, evaluation tools and serving stack under a license permitting commercial use. Alpamayo 2 Super directs another open model, with 34 billion parameters, toward autonomous-vehicle development. It combines a Cosmos 3 Super reasoner of 32 billion parameters with a diffusion action expert of 2 billion parameters. From surround video covering as many as seven cameras, it can produce future vehicle trajectories, reasoning traces, high-level actions, grounded scene answers and automatic labels. Weights are on Hugging Face and inference notebooks on GitHub under OpenMDW-1.1. Cosmos 3 supplies a broader base for robot policies, while Alpamayo 2 Super focuses on interpreting, labeling and evaluating driving footage and turning it into candidate vehicle behavior.[1], [2]
The content policy arrives at inference time
Mistral's Shieldstral is assigned a narrower job: classify whether a prompt, response, prompt-response pair or image complies with a supplied content policy. The model has 3 billion parameters and frames moderation as binary question answering. Criteria can address violence, age appropriateness or whether an assistant refused to answer; a product developer can describe the policy in plain language at inference time without retraining. This arrangement supplies rules written for a product's audience and context as model input instead of fixing one harm taxonomy in the weights. Because Shieldstral examines both text and images, it can consolidate some work that would otherwise require separate textual and visual guard models. Mistral published the weights on Hugging Face under Apache 2.0 and says the model can run on a single 16-gigabyte NVIDIA GPU. The company also says Shieldstral matches or exceeds open guard models up to seven times its size across four evaluation axes. That comparison is Mistral's own measurement, and the announcement presents charts rather than a numerical table. Multilingual coverage is a future priority, but supported languages are not listed. Present use is policy-adaptive text and image moderation; broader language coverage remains later work.[3]
A smaller body for a local agent
Liquid AI's LFM2.5-2.6B moves open weights into general-purpose agents that can operate away from a network service. The model has 2.6 billion parameters, and its context window was extended to 128,000 tokens during mid-training. Liquid AI says it runs in less than 2.5 gigabytes of memory and can generate 30 tokens per second on a phone. The model works with llama.cpp, MLX, vLLM, SGLang, ONNX and WebGPU; base and post-trained versions are available on Hugging Face. The accepted news card, however, states no license. Read together, the four releases do not describe one product category. Cosmos 3 supplies hardware-scaled foundations for robot-arm policies; Alpamayo 2 Super turns multi-camera driving data into trajectories and labels; Shieldstral accepts a product policy at inference time for safety classification; and LFM2.5-2.6B targets low-memory local agents. Mistral specifies Apache 2.0, and NVIDIA's two releases identify licenses allowing commercial use, while the Liquid AI announcement's news card gives no license. The group shows that access to weights is only one part of a release: its intended job, target hardware and distribution conditions must each be read separately.[4], [1], [2], [3]
