What Is FLUX 3: Black Forest Labs’ Multimodal Base Model

FLUX 3 is Black Forest Labs’ multimodal foundation model, released on 23 July 2026, that jointly learns images, video and audio in one architecture. The lab behind FLUX.2 spent early 2026 moving beyond still images, and FLUX 3 is the result: a system positioned as “visual intelligence,” not a single-purpose generator.

This overview covers what FLUX 3 is, when each part shipped, what you can use today, and what the image and video models do and cost. The API reference and the FLUX.2 comparison live in their own articles; this page links to both.

What FLUX 3 is

Black Forest Labs introduced FLUX 3 on 23 July 2026 as a multimodal foundation model trained through a method it calls Self-Flow, where one architecture jointly learns images, video and audio. The official framing is “visual intelligence”: a base that works across sight and sound rather than a single modality.

The roadmap goes past image and video: Black Forest Labs also lists action prediction for robotics, shipping first as FLUX 3 Action, an open-weight motion model, while image and video reach users via the API and Playground.

The rollout timeline

FLUX 3 shipped in pieces across three months; the open-weight backbone is still promised, not yet delivered.

FLUX 3 Dev, the open-weight backbone, was promised “in the coming weeks” at launch but had not shipped by 7 October 2026, with no date given.

What you can use today

Three channels carry FLUX 3 today, and they differ in price, limits and free tier.

Sign-up follows the FLUX.2 routine: register, confirm email, add credits, create a key. The FLUX 3 API guide covers the request, and the how-to-use guide walks the four-step sign-up.

What FLUX 3 Image can do

The image model is layout-aware, not a plain text-to-image box. Bounding-box layout uses a normalised 0–1000 grid, the element table appended to the prompt as JSON, with up to ten reference images taking tokens in order from ref_image_0.

Pixel-level local editing supports four operations — keep, move, add and remove. The vendor notes pixels outside the edited region “usually stay unchanged,” though elements may extend past the box. Resolution runs from 768 square to 4K (about 16MP, or 5456×3072 at 16:9), sides in multiples of 16, with 15 aspect ratios plus auto.

Black Forest Labs marks these as list prices as of 15:00 UTC on 8 October 2026, with a 50% launch discount as a limited promo. Prompts and reference images are not billed, and failed or blocked requests cost nothing.

What FLUX 3 Video can do

The video model is natively audiovisual: up to 20 seconds at 24fps with synchronised audio, where dialogue, lip sync, sound effects and ambience are generated together rather than dubbed on later.

Four modes cover the work — text-to-video, image-to-video with one to ten pinned frames, video-to-video continuation, draft-enhance — billed per second.

Black Forest Labs has published early human-preference figures for 10-second, 720p, audio-bearing text-to-video: 93% preference against Luma Ray 3.2, 77% against Runway Gen-4.5, and 52% each against Seedance 2.0 and Gemini Omni Flash. These are the vendor’s preliminary numbers from when the model was still training, not a settled ranking.

Worked examples: two copyable inputs

Two inputs you can paste today: one fixes a layout with a bounding box, the other composes a scene from reference tokens.

The first example places three elements on the 0–1000 grid, appended to the prompt as JSON: a backdrop, a bottle and a small nameplate.

[
  {
    "id": "backdrop_1",
    "bbox": [0, 0, 1000, 1000],
    "desc": "A flat seamless studio backdrop in pale grey, softly lit"
  },
  {
    "id": "bottle_1",
    "bbox": [300, 380, 820, 620],
    "desc": "A tall frosted glass bottle standing upright in the centre"
  },
  {
    "id": "plaque_1",
    "bbox": [60, 60, 240, 300],
    "desc": "A small blank nameplate in the upper left corner"
  }
]

Paste the JSON after your prompt and confirm the bottle stays inside its box; if the nameplate drifts, lower the bottom value.

The second example pulls two reference images into one sentence; ref_image_0 and ref_image_1 stand in for the files, in order.

A product scene with ref_image_0 standing on ref_image_1, lit by one soft key light, shallow depth of field, neutral studio backdrop

Use this when the subject and its surface must match supplied images; verify that both references appear and that the key light reads as a single source.

What Black Forest Labs has not published

Still unpublished: parameter counts, checkpoint sizes, image-side benchmark scores, and a firm FLUX 3 Dev date. The 50% launch discount ends on 8 October 2026, after which list prices return.

For the two-generation picture, the FLUX 3 vs FLUX.2 guide sets out what changed and what stayed. The FLUX topic hub collects every article on the family.

Frequently asked questions

Is FLUX 3 free? The Playground lets you try it, and fal gives five free text-to-video drafts per day to logged-in users. The image API has no free quota.

Can I run it locally? Not the full model yet — FLUX 3 Dev is still promised. FLUX 3 Action, a 7B open-weight model, is already available locally.

Is FLUX.2 still usable? Yes. FLUX.2 remains the current image line, and the comparison guide shows where each fits.

How much is one 4K image? The list price is $0.607, or $0.3035 during the 50% promo; prompts and references are not billed.

FLUX 3 is less a new image model than a multimodal base — and the parts you can use today are only the first slice of the plan.

Published

Oct 7, 2026

Information verified

Oct 7, 2026

By

AI Tool Blog Editorial Team