Astronex-World 1.0 Streams Open World-Model Video on a Single 48GB GPU

Astronex Robotics published Astronex-World 1.0 on September 17, 2026, an open video world model whose weights and code are downloadable now and whose causal variant streams video frame by frame at 832x480 on one Nvidia L20 GPU. The paper lists Xin Zhou of Astronex Robotics and Cong Miao of Nanjing University of Information Science and Technology as authors, and the model card reports 5.35 billion parameters under an Apache 2.0 licence.

A world model is a generator that predicts what happens next rather than rendering a finished clip, and Astronex-World takes that literally. Given a text prompt or a first frame, it predicts future visual states under frame-aligned camera trajectories, a 64-dimensional continuous action stream and an embodiment identifier, and it accepts text events inserted part-way through a rollout. It ships in two forms: a bidirectional model for full-context generation, and a causal model with block-causal attention and cross-block KV caching for generation that keeps going.

The numbers are in the model card. The causal model produces 832x480 video at 24 frames per second with 8-step sampling and, the authors say, streams in real time on a single L20 48GB card. Training ran in five stages, from bidirectional control adaptation to distribution-matching distillation, on two of those GPUs. On the WBench suite the release scores 73.5 on Navi and 70.0 on Full, which the card places above a 13.6B LongCat-Video at 69.9 and a distilled 14B Helios at 69.7, and within 0.9 points of the 22B LTX-2.3 at 70.9.

The paper's benchmark table is the authors' own and the VBench results it includes are explicitly partial, covering 240 of 6,220 text-to-video prompts. There is no third-party evaluation of Astronex-World yet, and no hosted product: it is a self-hosted release.

Running it is the interesting part for anyone with a workstation. The repository documents a memory-optimized inference path that frees its KV caches before decoding, dropping peak VRAM from 37.9GB to 23.2GB on identical output, which the project says brings it inside a 32GB consumer card.

This is a research release rather than a tool, so the comparison that matters is with the closed world models, and there it is the licence that separates them. Google's Genie 3 sits behind a subscription and a research preview; Astronex-World 1.0 is a download, and its 5.35B footprint is a fraction of the 22B open model it scores closest to.

Weights are on Hugging Face and code is on GitHub, both under Apache 2.0.