FLUX 3 Image open weights: release date and what to use now

There is no release date for FLUX 3 Image's open weights. As of 2026-10-03, press coverage of the Oct 1 launch says Black Forest Labs expects an open-weight version "in the coming weeks" (The Decoder) or "within weeks" (Tech Times), and neither article gives a date, a license, or a file size. Until the weights ship, FLUX 3 Image is reachable as a hosted model, for example on Replicate.

Sume does not list FLUX 3 Image in its image catalog today. What Sume serves from Black Forest Labs is FLUX.2 Pro and FLUX.2 Flex, so an open-weight release would not change what you can call on Sume until a FLUX 3 row appears in the catalog.

What has actually been said about the weights?

Both reports are news coverage, not a Black Forest Labs license page, so treat every line below as reported. The Decoder (Oct 2) says the open-weight version is expected in the coming weeks and that commercial licensing is available for on-premise deployment and fine-tuning. Tech Times (Oct 2) says the same about timing and adds that commercial licensing covers fine-tuning and self-hosted deployment.

Neither report names a license for the open weights. That matters because an earlier FLUX.2 family split its licensing by size, and the open release of a model is not the same thing as permission to sell what it makes. Wait for the license text on the model card before you plan a product around it.

• Reported: open weights "in the coming weeks", no date.
• Reported: commercial licensing for on-prem and fine-tuning exists separately.
• Not stated anywhere I could read: license terms, parameter count, hardware needs, or whether the 4K output and bounding-box editing ship in the open version.

What do the hosted FLUX 3 Image inputs look like today?

The Replicate page lists the inputs you would also expect from a self-hosted copy: a prompt, up to 10 reference images, an aspect ratio (auto or fixed ratios), a resolution of 768sq, 1k, 1.5k, 2k, or 4k, a grounding switch that is on by default, and an output format. Reference images must be between 256 by 256 pixels and 16 megapixels (read 2026-10-03).

Those are the same inputs worth designing around now, because a weights release usually keeps the hosted model's parameter names. If you write a thin wrapper today that maps your own fields onto them, swapping the backend later is a one-file change.

Weights or a hosted API: which fits a small team?

Open weights give you control of the data path and the cost curve at high volume, and they cost you a GPU, an inference stack, upgrades, and moderation. A hosted catalog gives you none of that control and none of the operations work. The table sets out what each route needs from you.

How do I find out the day FLUX 3 shows up on Sume?

Sume publishes its image catalog at GET /v1/images/models, so the cheapest watcher is a script that lists the ids and flags any with flux in the name. When a black-forest-labs/flux.3... row appears, its supported_parameters will tell you which of the inputs above it accepts, because Sume rejects any parameter a model does not list with 400 unsupported_parameter.

This script prints the FLUX rows in the current catalog. It needs only a Sume API key in SUME_API_KEY.

import os
import requests

r = requests.get(
    "https://api.sume.com/v1/images/models",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    timeout=30,
)
r.raise_for_status()
for m in r.json()["data"]:
    if "flux" in m["id"]:
        refs = m["supported_parameters"].get("input_references", {})
        print(m["id"], "max references:", refs.get("max"))

What should I build on in the meantime?

If you want Black Forest Labs output on Sume now, use FLUX.2 Pro or Flex; the FLUX 3 versus FLUX 2 note covers the difference. If you want the layout and region-edit features that are specific to FLUX 3, there is no like-for-like model in Sume's catalog; ChatGPT Image 2.5 has a mask_url edit that is the closest equivalent, and the bounding box note explains where it differs. For license reading habits, the FLUX.2 klein licensing post is a good template.

Re-check the Replicate page and Black Forest Labs' own docs before you commit: the release is described as imminent, and a license page is the only fact that counts. A plan that works whether or not the weights ship is the safest one: keep your prompts and reference images in your own storage, keep model ids in config, and treat the backend as swappable.