Black Forest Labs ships FLUX 3 — multimodal video, image, and robotics in one backbone
FLUX 3 expands Black Forest Labs’ stack into multimodal video-with-audio, image editing, and robotics in one backbone for creative and industrial work.
AI/TLDR lists FLUX 3 on July 23 as Black Forest Labs’ multimodal step: video with audio, image editing, and robotics on one backbone, plus a FLUX 3 x mimic variant said to drive factory robots at about 101 ms. Primary confirmation via the BFL site and Hugging Face at publish.
BFL made its name on stills. FLUX 3 is the company saying the backbone is now a control surface, not only a pretty decoder. That is a category jump. Treat it as a family announcement until each SKU has its own card.
Modalities in the FLUX 3 family
Tracker copy groups video-plus-audio generation, image editing, and robotics under one family name. That is a lot of product for one brand. The Models job is to keep the SKUs straight: a video model that can also edit stills is not automatically a robot policy. If BFL shipped separate checkpoints under one family, say so from the blog.
Audio-on-video is the consumer hook. Robotics is the industrial hook. Do not let a launch film imply both are one download. Image editing is the continuity product — the thing existing FLUX users will actually try first.
Robotics spin-off latency claims
About 101 ms for FLUX 3 x mimic on factory robots is the number trackers repeated. One hundred milliseconds is in the range where a closed loop can feel alive if the rest of the stack cooperates. It is also easy to measure in a cherry-picked cell.
Ask the BFL post what 101 ms includes: perception, policy, or just the model forward pass. A 101 ms network without camera and controller time is a lab number. A 101 ms end-to-end cycle on a real line is a product number. If the post is silent, assume forward pass.
Open vs API access
BFL has historically mixed open FLUX weights with paid APIs. FLUX 3’s access model will decide who prototypes. If video and robotics stay API-only while a stills distill is open, the “one backbone” story is a marketing umbrella. Read the license on each card. An API-only robot policy is a vendor relationship, not a community weight.
Who should prototype on FLUX 3
Video teams that already like FLUX stills and want audio in the same family. Robot integrators who can measure 101 ms themselves on hardware they already own. Image editors who want the next FLUX in-painting step. Not teams that need a published humanoid safety stack — that is DeepMind’s July 30 lane, not BFL’s stills heritage.
FLUX 3 Black Forest Labs is a July 23 multimodal claim. Believe the BFL hero over the tracker if they diverge. Measure the robot loop before you rewrite a cell around 101 ms.
