![Banseog [ physical ai ]](/assets/brand/banseog-physical-ai-on-dark.svg?v=20261007){kind=small}
Banseog note · 2026-10-26
A lab extending from image generation models to robot hands
Domain expansion — the stage of linking image and video models to robot action prediction
- According to a July 2026 company press release, a team of about 100 people works at its headquarters in Freiburg, Germany and in San Francisco.
- Founded by the researchers who created Latent Diffusion and Stable Diffusion, it says it has raised more than USD 450 million from investors including Salesforce Ventures, NVIDIA and Adobe Ventures.
- FLUX 3, released on July 23, 2026, is a multimodal foundation model that learns image, video, audio and action prediction together in a single architecture; the company says it cut the data a robot needs to learn a new manipulation task from more than 30 hours to about 30 minutes.
Decisions within a year
- FLUX 3 and FLUX-mimic release (July 23, 2026): a video-action model built with robotics company mimic robotics; the company says it is used at Audi production sites for parts kitting, inserting electronic control units and handling flexible materials such as cables and thread.
- FLUX 3 Action release (September 23, 2026): a 7-billion-parameter open-weight world-action model that predicts robot actions and future video frames together; it recorded a 42.2% success rate on the RoboLab-120 benchmark and was released on Hugging Face and GitHub.
- FLUX 3 Image Klein release (October 15, 2026): it released the weights of a 5.6-billion-parameter image generation and editing model and allowed free commercial use for individuals and companies with annual revenue under USD 5 million.
People they will look for
- Robot learning researcher (world-action models) — research to raise FLUX 3 Action's manipulation performance on real robot arms
- Robot evaluation and deployment engineer — validating models with partner robot companies and at production sites, and putting them on edge devices (Jetson and others)
- Large-scale training infrastructure engineer — making training and inference more efficient for multimodal models that learn image, video and action together
For hiring managers
This company hires at its offices in Germany and the United States, so Korean talent would need to relocate. In Korea, large companies' AI labs, humanoid and robot startups and the Korean subsidiaries of global AI companies look for the same robot learning and generative model researchers, so now, just after it released open-weight models, is the time to meet researchers in this field.
For job seekers
Research or implementation experience connecting video generation or diffusion models to robot manipulation policy learning is worth the most at this company right now. Work locations are Freiburg or San Francisco.
An analysis written from public sources. Sources: Yahoo Finance/GlobeNewswire 2026-07-23 (FLUX 3 release, team of about 100, more than USD 450 million raised, 30 hours to 30 minutes), BFL blog 2026-07-23 (FLUX 3 x mimic, Audi sites), BFL blog 2026-09-23 (FLUX 3 Action, RoboLab-120 42.2%), AI Times 2026-10-17 (FLUX 3 Image Klein), BFL careers page (Freiburg, San Francisco).
Roles hiring now
Skills asked for most
Open jobs
Banseog checks official company careers pages every day. A job is marked closed only when the official page returns a removal (404 or 410).