Black Forest Labs Open-Weights FLUX 3 Action, a 7B Robotics Model
TL;DR
- BFL's self-reported 42.92% RoboLab-120 score was not listed on NVIDIA's public leaderboard at announcement, leaving the claim unverified.
- FLUX 3 Action fits on a 24GB consumer GPU via FP8 quantization, removing the data-center hardware requirement faced by its 16B rivals.
- Third-party Positronic Robotics testing on a Franka arm logged 93.3% task success, the only independent hardware result available at launch.
Black Forest Labs has open-weighted FLUX 3 Action, a 7-billion-parameter world-action model, and reports it scores 42.92% on NVIDIA's RoboLab-120 benchmark, 6.1 points above the previous leader, NVIDIA's own 16-billion-parameter Cosmos3-Nano-Policy. Per VentureBeat, the model runs 1.43 times faster than Cosmos while using 44% of its parameters. The prior top score of 39.0% belonged to OASIS WAM.
The benchmark is NVIDIA's, which is what makes the result awkward. VentureBeat describes RoboLab-120 as "an NVIDIA-developed simulation benchmark intended to test general-purpose robotic policies across 120 tasks, spanning visual understanding, relational reasoning and procedural skills."
FLUX 3 Action ingests camera frames, robot state and text instructions, and predicts the next 32 actions plus how the scene will evolve visually. BFL says it can be fine-tuned from roughly 200 teleoperated episodes. The release ships weights, code, a fine-tuning recipe, benchmarks, reproducible examples and integration with Hugging Face's LeRobot framework. Missing from the release: an Action API, and a specific commercial license for the weights.
BFL has prior industrial ground via FLUX-mimic, built with Swiss startup mimic robotics, which the article reports was tested in Audi production environments for industrial manipulation tasks. Drone-flight and Doom-gameplay demonstrations are characterized as "early experiments rather than production capabilities."
BFL closed a $300 million Series B at a $3.25 billion valuation in December 2025, taking total funding above $450 million. The release lands in a busy open-robotics week that also included Alphabet's Intrinsic open-sourcing its Intrinsic Core stack at ROSCon.
What others are reporting
-
Black Forest Labs Read →
Publishes full training recipe with pretraining/midtraining data splits, third-party Positronic Robotics validation (93.3% on Franka arm), and cost-per-success breakdown.
Our single step 7B checkpoint outperforms every other open policy on RoboLab at 38.3% ± 0.38 compared to 36.8% for Cosmos 3 Nano.
-
Hugging Face (Black Forest Labs) Read →
Model card specifies six deployment variants, 24GB minimum GPU with FP8 quantization, and explicit safety warning that joint position outputs are unbounded.
-
AlphaSignal Read →
Frames the story as hybrid-system economics: FLUX 3 Action paired with GPT-6 Astra cuts per-task inference cost 29% vs pure reasoning-model approaches.
A capable low-level policy reduces the number and duration of reasoning-model calls.
-
Superpower Daily Read →
Flags that the claimed benchmark score had not appeared on NVIDIA's public leaderboard at publication, and that promised model weights were unavailable.
The robot observes its surroundings again and repeats the process. The aim is to make each set of movements responsive to what the camera sees next.
-
Crypto Briefing Read →
Applies the sharpest skeptical lens: notes no public confirmation of the model on RoboLab-120 and calls the leaderboard-topping claim unvalidated.
The 7-billion-parameter open-weight model claims to outperform larger competitors in robotic manipulation tasks while running on consumer hardware.
Originally reported by venturebeat.com
Read the original article →Original headline: Black Forest Labs Open-Weights FLUX 3 Action, a 7B World-Action Model Topping Nvidia RoboLab-120