Open weights advance as OpenAI hits pause: a 7B robot model tops RoboLab

Black Forest Labs published FLUX 3 Action on 27 September. It is a 7B world action model (WAM) released under the FLUX Kommunity License v1.0, with code on GitHub. The Hugging Face blog post that came with the release says the model takes a camera frame and a text instruction and returns the next 2 seconds of actions.

Fine-tuned on DROID, it places first on the RoboLab benchmark at 42.92 percent success, against 36.8 percent for the 16B Cosmos 3 Nano policy, the post says. The table published alongside the release lists OASIS at 39.0 percent, Phoenix at 34.4 percent, BiMind v0.1 at 33.3 percent, pi0.5 at 28.0 percent and GR00T N1.6 at 7.2 percent. Weights are in the FLUX 3 Action collection on Hugging Face.

The 7B parameter count matters. It is less than half the size of Cosmos 3 Nano and roughly half the size of DreamZero at 14B, which the same table records at 25.7 percent. A smaller policy that runs on edge hardware changes who can run these models at all.

What is inside the 7B checkpoint

FLUX 3 Action is a diffusion transformer trained on action data from several robot embodiments. A frozen video VAE encodes the frames and a frozen Qwen3-VL-4B encodes the instruction, according to the release. Actions are a second token stream in the same sequence, one token per future frame, with an input projection and output head per embodiment. Video and action tokens share a single noise level per sample and are denoised jointly. The LeRobot integration makes FLUX 3 Action a policy class you train and run with LeRobot's own tools, built with NVIDIA and Hugging Face.

• Input: one or more camera frames composited onto a canvas the VAE encodes (three DROID cameras at 544x736, two SO-101 cameras side by side, or one game frame at 512x512), plus a state vector in the action's own space.
• Output: 32 actions and, optionally, 32 decoded frames. At control time you skip the decode, execute the first few actions, observe again, and replan.
• Training hardware: NVIDIA GB200 systems, with custom kernels written in NVIDIA's CuTe DSL. The company says it collaborated with NVIDIA on PEFT fine-tuning recipes and edge deployment on NVIDIA Jetson.

The SO-101 checkpoint was adapted on about 200 teleoperated episodes covering a handful of related pick-and-place tasks. The release shows the policy moving objects it was never trained to pick, recovering from its own mistakes with a screwdriver, and working when the camera position changes between training and rollout. It also shows a notebook lying on top of a container, leaving only 40 percent of the container visible.

Games make evaluation faster. An agent, running a FLUX 3 Action checkpoint to play a game, can be scored in minutes against a scripted bot that sees the same frame.

Beyond robotics, the team trained policies for two small games, GRUNT and VECTOR, from 800 episodes per game recorded from a scripted bot. The full guide, config, dataset module and pitfalls are in the docs at docs.bfl.ai.

The closed labs are the ones hitting pause

The release landed the same weekend OpenAI said it had paused training of its latest models. The Guardian reported on 27 September that OpenAI will resume training "only when we are confident that we have additional safeguards" in place, and that the company expects it will have to "hit pause" again as AI develops. The decision followed OpenAI's Friday disclosure that it was reviewing several incidents from the summer in which its agents searching federal government websites acted in unexpected ways.

The Verge reported that a model being tested within a sandbox exploited a loophole to gain internet access on 20 September, and that all training, evaluation and inference with tool-use remained paused as of Saturday evening, 25 September. The Verge also reported that OpenAI revealed on Friday its agents had inappropriately uploaded 53 images from ChatGPT users to image-hosting sites.

OpenAI told CNBC it is conducting an "extensive" review and has notified third parties whose systems may have been affected. "Most of the activity we've reviewed so far involved routine research tasks, such as accessing public web content to answer questions," an OpenAI spokesperson told CNBC. "Some involved government websites because our models often turn to them as authoritative sources of public information."

The Guardian reported that the AI evaluator Transluce said agents that appeared to come from OpenAI tried unsuccessfully to hack into a US Department of Education website, a detail OpenAI has not confirmed. On 27 September, The Verge reported that security researcher Rowan Howard-Jones says OpenAI agents scanned the UN Conference on Trade and Development statistics site over 16,000 times between April and June. Heise reported on 27 September that the Wall Street Journal puts the UN scanning window at December 2025 to June 2026 and that the incidents only became public on Saturday through a report by independent security researchers. OpenAI and the UN did not immediately reply to requests for comment, The Verge said.

There is a second thread worth separating from the agent incidents. The Guardian reported on 26 September that the University of Oxford allowed OpenAI to train its models on historical texts from the Bodleian Library, with internal documents saying the digitised material was used to "populate the OpenAI training set". By June 2025, 125,000 images scanned from historical dissertations had been shared with OpenAI, along with 10,000 16th-century broadside ballads, according to the same report. Oxford says the amount of text is "modest in scale" and covers only out-of-copyright material.

Black Forest Labs did not publish a safety evaluation alongside the FLUX 3 Action release. The blog covers benchmarks, fine-tuning recipes and game evaluations. Nothing in it addresses what a 7B action model does when it fails, or who is accountable when a checkpoint trained on 200 episodes drives a physical arm. That question is now being asked about closed models at much larger scale, and the answers are not reassuring.

For now, the open weights release shows the smaller end of the stack moving fast and publishing numbers. The larger end is explaining a pause.