Black Forest Labs Releases FLUX 3 Action Multimodal World Action Model

Introduction

On September 22, 2026, Black Forest Labs (BFL) officially bridged the gap between digital intelligence and physical agency with the release of FLUX 3 Action. As the first open-weights entry in the highly anticipated FLUX 3 family, this model represents a fundamental shift from generative media to generative physical intelligence.

While BFL has previously dominated the latent diffusion space with high-fidelity image and video models, FLUX 3 Action introduces a 'World Action Model' (WAM) architecture. This release signals BFL's commitment to a unified multimodal frontier, where visual understanding directly translates into actionable physical commands.

• Release Date: September 22, 2026
• Model Type: Multimodal World Action Model (WAM)
• Parameter Count: 7B
• License: Open-weights (Non-gated)

Key Features & Architecture

FLUX 3 Action is built on a sophisticated 7-billion parameter architecture that integrates visual perception with temporal action prediction. Unlike traditional robotics models that rely on rigid heuristic controllers, FLUX 3 Action utilizes a unified transformer-based approach to process multimodal inputs in a single forward pass.

The core mechanism involves taking a single camera frame combined with a natural language instruction. The model then predicts the optimal sequence of actions for the next 2 seconds of operation. This is achieved through a dual-stage process: a video generation component predicts how the world state will evolve, while an Inverse Dynamics Model (IDM) translates those predicted visual changes into precise motor control signals.

• Input Modalities: RGB Camera Frames + Text Instructions
• Prediction Horizon: 2-second continuous action sequences
• Architecture: Unified World Action Model (WAM) with Inverse Dynamics
• Deployment Variants: Base checkpoints, Droid-specific, and SO-101 robot variants

Performance & Benchmarks

In early benchmarking, FLUX 3 Action has demonstrated industry-leading efficiency. Despite its relatively compact 7B parameter footprint, it consistently outperforms much larger proprietary models in manipulation tasks and spatial reasoning. The model's ability to generalize across different object geometries and lighting conditions sets a new baseline for open-weights robotics.

The model was evaluated using standardized robotics benchmarks focusing on success rates in pick-and-place, tool use, and obstacle avoidance. Specifically, the 'Action-Success-Rate' (ASR) metric showed that FLUX 3 Action achieves state-of-the-art performance on the Droid and SO-101 hardware platforms, effectively doubling the reliability of previous open-source iterations in unstructured environments.