Black Forest Labs tops Nvidia's own robot benchmark days before Germany's Nvidia summit

A German lab takes Nvidia's benchmark

Black Forest LabsPVT~$4B — Black Forest Labs, private, latest valuation ~$4B, the Freiburg and San Francisco lab behind the FLUX image generators, has released FLUX 3 Action↗, a 7-billion-parameter open-weight model that turns camera feeds and written instructions into robot motion ¹²³. It completes 42.9 percent of 1,200 trials on RoboLab-120, the simulation benchmark Nvidia▼−1.70% — Nvidia, down 1.70 percent today presented in July and runs itself, for first place ¹². Handelsblatt reported on October 15 that Berlin is set to unveil a physical-AI agreement with Nvidia at Monday's Invest in Germany summit↗, which Nvidia chief executive Jensen Huang and Chancellor Friedrich Merz are due to attend ⁴⁵.

Nvidia's own entries trail on its leaderboard: its 16-billion-parameter Cosmos3-Nano-Policy scores 36.8 percent and its GR00T N1.6 robot model just 7.2 ¹². World-action models, the video-trained class FLUX 3 Action belongs to, hold four of the benchmark's top five spots ¹.

On a real arm, the lead narrows to one attempt

In blind tests the lab arranged with Positronic Robotics, an outside evaluation firm, a Franka robot arm on FLUX 3 Action completed 28 of 30 attempts across ten tasks, against 27 for Cosmos 3 Nano and 13 for Physical Intelligence's π0.5 ¹². The operator did not know which model drove which attempt ².

Germany's answer is a deal with Nvidia

The reported deal carries hundreds of millions of euros of state money for AI compute purchases and model training, with Nvidia contributing engineers ⁴⁵. Roughly a dozen German companies would take part, Siemens, Audi and Black Forest Labs among them ⁴.

Reuters reports the agreement would be built around SOOFI, the Sovereign Open Source Foundation Models consortium launched in September 2025 and funded with €20 million by Germany's economics ministry↗ ⁵⁶. SOOFI's first model was trained on Deutsche Telekom's Industrial AI Cloud in Munich, a site expanded with Nvidia for roughly €1 billion ⁵⁷. No comment was available from the German government ⁶.

Open weights, Nvidia plumbing

Robin Rombach, who leads the lab, argued in the French daily L'Opinion on October 10↗ that open model weights strengthen Europe's strategic independence ⁸. The lab's own credits complicate that pitch: FLUX 3 Action was trained on Nvidia GB200 systems, Nvidia helped move it to Jetson edge computers, and BFL thanks Nvidia's Isaac Lab and RoboLab teams for "the basis for most of our evaluations" ¹².

The lab, about 100 people, is valued at $3.25 billion on more than $450 million raised, with Nvidia among the investors per DataNorth, an industry tracker ³. Free commercial use of the weights stops at $5 million of annual revenue ¹.

The record awaits real robots

Open weights are the strongest form of the sovereignty case: a lab with a low-cost SO-101 arm and about 200 demonstrations can fine-tune the top-ranked model itself ¹. RoboArena, the independent ladder that scores models on real arms, has listed none of RoboLab's top five↗, and on the benchmark's hardest tasks two rivals outrank FLUX 3 Action ¹.

On the record so far, Europe's independence project runs through Nvidia, which supplies the chips, the benchmark and the evaluation code, and, per one tracker, a stake in the winning lab ¹²³. Rival silicon suppliers lose the order if Monday's compute budget lands there, and labs above the $5 million line pay Black Forest Labs for a license either way ¹⁵. Watch Monday for the named compute supplier and participant list, and RoboArena for FLUX 3 Action's first real-robot entry ¹⁵.