Video models became robot controllers, and the latency fell from 800 ms to 62 ms
Runway's Praxis-1 predicts video and actions together. Reactor cut its latency to 62 ms. The same week brought faster streaming video and sharper image editing.
- Praxis-1 achieved a 46.2% success rate on the RoboLab-120 leaderboard across 1,200 episodes, ahead of FLUX 3 Action at 42.9% and Cosmos3 Nano at 36.8%. Self-reported m:…9277
- Reactor reduced the inference latency of the Praxis-1 model from 800 ms to 62 ms, and offers a factory-tier service for Praxis-1 with a 62 ms latency budget. Self-reported m:…4806
- Praxis-1 reached 80.0% success on SimplerEnv when fine-tuned on the RT-1 robot, against 52.4% for the original RT-1 policy. Self-reported m:…9277
- APEX reduces peak allocated GPU memory per GPU by 4.4% compared to DMD, and reduces mean iteration time by 9.2% compared to DMD. Self-reported m:…2397
A video model that acts
Runway has released Praxis-1 under an Apache 2.0 license, and the model is not a video generator in the usual sense. Praxis-1 is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions simultaneously, which makes it a world action model: it imagines what happens next and decides what to do about it in the same pass. Praxis-1 supports full fine-tuning and LoRA, and runs on NVIDIA hardware including H100, B200, RTX 5090, RTX PRO 6000, and Jetson Thor. m:…9277
The benchmark numbers put it ahead of the open-weight field. Praxis-1 achieved a 46.2% success rate on the RoboLab-120 leaderboard across 1,200 episodes, ahead of FLUX 3 Action at 42.9% and Cosmos3 Nano at 36.8%. Praxis-1 reached 80.0% success on SimplerEnv when fine-tuned on the RT-1 robot, against 52.4% for the original RT-1 policy. Developers can use NeMo Automodel tutorials to fine-tune Praxis-1 for domain-specific use cases, and Runway is a founding member of the Cosmos Coalition, which is building shared foundations for physical AI. m:…9277
Latency is the product
A robot controller that takes 800 ms to decide is a robot that acts on stale information, and Reactor's optimization is the week's most practical result. Reactor reduced the inference latency of the Praxis-1 model from 800 ms to 62 ms, and offers a factory-tier service for Praxis-1 with a 62 ms latency budget. Reactor's standard offering for Praxis-1 maintains a p99 latency under 270 ms, and a single Reactor server can support sixteen robots while maintaining a p99 latency under 250 ms. m:…4806
The optimization was subtraction, not addition. Reactor optimized the Praxis-1 inference by removing work that does not reach the final action output, and uses CUDA graphs to remove GPU wait times on Python execution. Praxis-1 uses a chunk-causal window where one pass receives d chunks and denoises one chunk, and generates video and actions using rectified flow. Reactor's early VAE encoding of camera frames reduces call-time latency by approximately 19-20 ms. Robotics success rates decrease when actions arrive late, which is why the latency number is the headline, not the benchmark score. m:…4806
The same week in generation
The generation side kept moving too. Seedream 5.5 is available in ComfyUI via Partner Node, and is a multimodal model that generates from text and edits from images. Seedream 5.5 supports multi-turn editing where targeted changes are made while the rest of the frame remains unchanged, and supports localized control for changing an object's color, shape, or material without regenerating the whole image. Seedream 5.5 generates RGBA natively with support for transparent and semi-transparent backgrounds, and can extract a subject from an existing image onto a transparent layer. m:…5421
Seedream 5.5 supports text rendering and structured layouts for posters and infographics in a single pass, and provides improved photographic detail, including natural lighting, material rendering, and skin texture. Partner Node in ComfyUI run on credits, which is the pricing detail that matters for anyone building a workflow on it. The model is aimed at the working end of image generation: editing, compositing, and posters, not just first drafts. m:…5421
The research side of streaming video got faster as well. APEX replaces a separate fake-score network with condition shifting to estimate the generated distribution's velocity field, and uses condition shifting to allow the generator and fake estimator to share one network. APEX achieves an instruction following score of 3.856 compared to 3.831 for S2-Avatar, and a visual aesthetics score of 0.641898 compared to 0.632606 for S2-Avatar. APEX reduces peak allocated GPU memory per GPU by 4.4% compared to DMD, and reduces mean iteration time by 9.2% compared to DMD. This is a preprint, so the results are the authors' own and not yet peer-reviewed. m:…2397
Why it matters
The thread across the week is that video models are leaving the studio and entering the control loop. A model that predicts video and actions together, served at 62 ms, is a robot's eyes and hands as a single system. A streaming video model that shares one network between generator and estimator is a cheaper way to keep the picture moving. And an image editor that keeps the rest of the frame unchanged while you change one object is the tool a working designer actually needs. The generation race is becoming a latency and control race, and the winners will be the models that act in time. m:…9277 m:…4806 m:…2397 m:…5421
What is still unresolved
- Whether Praxis-1's RoboLab-120 lead holds on robots beyond the ones tested.
- Whether Reactor's 62 ms factory tier is available beyond its current robotics partners.
- Whether APEX's efficiency gains hold at production scale, given it is a preprint.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.