# Flux (text-to-image model)

> Source: https://aiwiki.ai/wiki/flux
> Updated: 2026-10-18
> Fact-checked: 2026-10-18
> Categories: Diffusion Models, Generative AI, Image Generation, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Flux (text-to-image model)." aiwiki.ai, 18 Oct 2026. https://aiwiki.ai/wiki/flux
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Flux** (styled FLUX) is a family of image generation models, and since 2026 multimodal video, audio and action models, developed by [Black Forest Labs](https://aiwiki.ai/wiki/black_forest_labs) (BFL), the Freiburg-headquartered startup founded by researchers behind [Stable Diffusion](https://aiwiki.ai/wiki/stable_diffusion) [33][42]. The first generation, [FLUX.1](https://aiwiki.ai/wiki/flux_1), was released on August 1, 2024. All public FLUX.1 models share a hybrid [transformer](https://aiwiki.ai/wiki/transformers) architecture of about 12 billion parameters trained with [flow matching](https://aiwiki.ai/wiki/flow_matching), and BFL claimed at launch that FLUX.1 [pro] and [dev] surpassed Midjourney v6.0, DALL-E 3 (HD) and SD3-Ultra on visual quality, prompt following, typography and output diversity [33]. The second-generation [FLUX.2](https://aiwiki.ai/wiki/flux_2) family, announced on November 25, 2025, couples a Mistral-3 24B vision-[language model](https://aiwiki.ai/wiki/language_model) with a [rectified flow](https://aiwiki.ai/wiki/rectified_flow) transformer, with the open-weight FLUX.2 [dev] at 32 billion parameters; Black Forest Labs describes it as "designed for real-world creative workflows, not just demos or party tricks" [8][46]. The third generation, FLUX 3, was announced on July 23, 2026 as a single model trained jointly on images, video and audio; its video product ([FLUX 3 Video](https://aiwiki.ai/wiki/flux_3_video)) became generally available on August 4, 2026, the image product ([FLUX 3 Image](https://aiwiki.ai/wiki/flux_3_image)) on October 1, 2026, and a 5.6-billion-parameter open-weight FLUX 3 Image Klein followed on October 15, 2026 [43][48][50][52].

This page is a hub for the whole family. Detailed coverage of each generation is on the [FLUX.1](https://aiwiki.ai/wiki/flux_1), [FLUX.2](https://aiwiki.ai/wiki/flux_2), [FLUX 3 Image](https://aiwiki.ai/wiki/flux_3_image) and [FLUX 3 Video](https://aiwiki.ai/wiki/flux_3_video) pages.

Flux models have been widely adopted across the AI image generation ecosystem. In August 2024, [xAI](https://aiwiki.ai/wiki/xai) used FLUX.1 to power image generation in the [Grok](https://aiwiki.ai/wiki/grok) [chatbot](https://aiwiki.ai/wiki/chatbot) on X [2][58]. [Mistral AI](https://aiwiki.ai/wiki/mistral_ai) added Flux Pro image generation to [Le Chat](https://aiwiki.ai/wiki/le_chat) in November 2024 [60], Freepik serves FLUX through its AI Suite at a scale of millions of inference requests per day [11], and [Adobe](https://aiwiki.ai/wiki/adobe) made FLUX.1 Kontext [pro] available in the Photoshop (beta) Generative Fill tool from September 25, 2025 [41]. In September 2025, Bloomberg reported that Meta had signed a multi-year contract to use BFL's image generation technology, worth $35 million in the first year and $105 million in the second [15][57]. On December 1, 2025, Black Forest Labs announced a $300 million Series B at a $3.25 billion post-money valuation [3][42].

## What is Black Forest Labs?

### Founding

Black Forest Labs was announced on August 1, 2024, alongside the first FLUX.1 models [33]. It was co-founded by [Robin Rombach](https://aiwiki.ai/wiki/robin_rombach), Andreas Blattmann and Patrick Esser, former researchers who helped create [Stability AI](https://aiwiki.ai/wiki/stability_ai)'s Stable Diffusion models [3][30]. Dominik Lorenz is also named as part of the founding team in TechCrunch's reporting [5]. According to investor Northzone, the three co-founders spent years in university research labs including LMU Munich and Heidelberg University, and their papers on generative image and video models have collectively been cited more than 120,000 times [30]. In its launch post, the company credited its team with VQGAN, Latent Diffusion, Stable Diffusion XL, Stable Video Diffusion, Rectified Flow Transformers and Adversarial Diffusion Distillation [33].

Robin Rombach is the lead author of "High-Resolution Image Synthesis with Latent Diffusion Models" (first posted to arXiv in December 2021), commonly known as the latent diffusion paper, which performs the diffusion process in the compressed latent space of a pretrained autoencoder rather than directly in pixel space. This reduced the computational cost of diffusion-based image synthesis and became the basis for Stable Diffusion [66].

Andreas Blattmann, Dominik Lorenz and Patrick Esser were co-authors of the same paper, alongside Bjorn Ommer [66]. Esser was also the first author of "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis" (2024), the Stable Diffusion 3 paper, on which Blattmann, Lorenz and Rombach were co-authors [68]. BFL lists Freiburg as its headquarters and also has a base in San Francisco [8][42].

### Funding

| Round | Date | Amount | Valuation | Lead Investors |
|---|---|---|---|---|
| Seed | Announced August 1, 2024 | $31M | ~$150M (post-money, per TechCrunch) | [Andreessen Horowitz](https://aiwiki.ai/wiki/andreessen_horowitz) (a16z) |
| Series A | Not formally announced | Not disclosed | Not disclosed | Not disclosed |
| Series B | December 1, 2025 | $300M | $3.25B (post-money) | Salesforce Ventures, Anjney Midha (AMP) |

The seed round, announced simultaneously with the launch of FLUX.1, was led by [Andreessen Horowitz](https://aiwiki.ai/wiki/andreessen_horowitz), with angel investors Brendan Iribe, Michael Ovitz, Garry Tan, Timo Aila and Vladlen Koltun, and follow-up investments from [General Catalyst](https://aiwiki.ai/wiki/general_catalyst) and MätchVC [33][4]. Michael Ovitz and Matthias Bethge were named to the company's advisory board at the same time [33]. In September 2024, TechCrunch reported from multiple sources that the company was closing a new round of about $100 million at a $1 billion valuation, a large step up from the seed round's $150 million post-money valuation, while cautioning that the deal might not be final [5].

BFL did not issue a public announcement for its Series A. Northzone later wrote that it had invested "at Series A" [30].

In December 2025, Black Forest Labs announced a $300 million Series B at a $3.25 billion post-money valuation, co-led by Salesforce Ventures and Anjney Midha's AMP. BFL said it was deepening its work with existing investors a16z, [NVIDIA](https://aiwiki.ai/wiki/nvidia), Northzone, Creandum, Earlybird VC, BroadLight Capital and General Catalyst, and welcoming Temasek, Bain Capital Ventures, Air Street Capital, Visionaries Club, [Canva](https://aiwiki.ai/wiki/canva) and Figma Ventures [42][3]. Crunchbase News put the company's total funding at $450 million after the round, per Crunchbase data [56].

### Commercial Partnerships

In September 2025, Trending Topics, citing Bloomberg, reported that Meta had signed a multi-year contract to use BFL's AI image generation tools, committing to "$35 million in the first year and $105 million in the second year"; a BFL spokesperson declined to comment on "current or potential partnerships". The same report said Bloomberg put the startup's annual recurring revenue at $96.3 million [57][15]. Meta's chief AI officer Alexandr Wang said that month that Meta had partnered with Midjourney and Black Forest Labs for the early version of its Vibes video feed [61]. In its Series B announcement, BFL named Adobe, Canva, Meta and Microsoft among partners building on its models, and TechCrunch listed Adobe, fal.ai, Picsart, ElevenLabs, VSCO and Vercel among companies using them [42][3].

## Model generations at a glance

| Generation | First release | Core design | Open-weight members | Detail page |
|---|---|---|---|---|
| FLUX.1 | August 1, 2024 | ~12B hybrid multimodal and parallel diffusion transformer, flow matching, CLIP and T5 text encoders | [schnell], [dev], Tools [dev], Kontext [dev], Krea [dev] | [FLUX.1](https://aiwiki.ai/wiki/flux_1) |
| FLUX.2 | November 25, 2025 | Latent flow matching; Mistral-3 24B VLM plus rectified flow transformer (32B [dev]); new VAE | [dev], [klein] 4B and 9B (plus Base and KV variants), VAE | [FLUX.2](https://aiwiki.ai/wiki/flux_2) |
| FLUX 3 | July 23, 2026 (announcement) | One multimodal flow model trained jointly on images, video and audio, built on BFL's Self-Flow approach | FLUX 3 Action (7B), FLUX 3 Image Klein (5.6B) | [FLUX 3 Image](https://aiwiki.ai/wiki/flux_3_image), [FLUX 3 Video](https://aiwiki.ai/wiki/flux_3_video) |

Sources: [33][8][48][51][52].

## FLUX.1 Models

The initial FLUX.1 release on August 1, 2024 comprised three model variants, each targeting different use cases and operating under different licensing terms [33].

| Model | Parameters | Training note | License | Availability | Target Use Case |
|---|---|---|---|---|---|
| [FLUX.1](https://aiwiki.ai/wiki/flux_1) [schnell] | 12B | Latent adversarial diffusion distillation, 1-4 steps | Apache 2.0 | [Open weights](https://aiwiki.ai/wiki/open_weights) ([Hugging Face](https://aiwiki.ai/wiki/hugging_face)) | Local development, personal use |
| [FLUX.1](https://aiwiki.ai/wiki/flux_1) [dev] | 12B | Guidance-distilled from FLUX.1 [pro] | FLUX.1 [dev] Non-Commercial License | [Open weights](https://aiwiki.ai/wiki/open_weights) ([Hugging Face](https://aiwiki.ai/wiki/hugging_face)) | Research, non-commercial use |
| [FLUX.1](https://aiwiki.ai/wiki/flux_1) [pro] | 12B | Not disclosed | Proprietary | API (BFL, Replicate, fal.ai) | Professional/commercial use |

All FLUX.1 variants support aspect ratios and resolutions between 0.1 and 2.0 megapixels [33].

### FLUX.1 [schnell]

Schnell (German for "fast") is BFL's fastest FLUX.1 model, described as tailored for local development and personal use. Its model card describes it as a 12-billion-parameter rectified flow transformer trained with latent adversarial diffusion distillation so that it can generate images in 1 to 4 steps. It is released under the Apache 2.0 license, allowing personal, scientific and commercial use [33][34]. BFL called it "the most advanced few-step model to date" at launch [33].

### FLUX.1 [dev]

The dev variant is an open-weight, guidance-distilled model "directly distilled from FLUX.1 [pro]", which BFL said obtains similar quality and prompt adherence while being more efficient than a standard model of the same size [33]. Its weights are released under the FLUX.1 [dev] Non-Commercial License; the model card states that generated outputs can be used for personal, scientific and commercial purposes as described in that license [35]. Since June 2025, businesses can buy commercial licenses for FLUX.1 [dev] and BFL's other open-weight models through a self-serve licensing portal [36]. In its FLUX.2 announcement, BFL called FLUX.1 [dev] "the most popular open image model globally" [8].

### FLUX.1 [pro]

The professional variant was available only through BFL's API and partner platforms such as Replicate and fal.ai, and its weights were not released [33]. BFL deprecated the FLUX.1 [pro] 1.0 API endpoints (flux-pro-1.0 and its depth and canny variants) on October 31, 2025, directing users to FLUX1.1 [pro] or FLUX.1 Kontext [pro] [43].

## FLUX1.1 [pro] and Ultra

### FLUX1.1 [pro]

Released on October 2, 2024 alongside the general availability of the beta BFL API, FLUX1.1 [pro] was a major upgrade that, in BFL's words, "provides six times faster generation than its predecessor FLUX.1 [pro]" while also improving image quality, prompt adherence and diversity. In the same post, BFL said it had updated FLUX.1 [pro] to run twice as fast, and that FLUX1.1 [pro] was three times faster than the then-current FLUX.1 [pro] [6]. The model was tested in the [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) image arena under the codename "blueberry" and, as of October 1, 2024, held the highest Elo score on that leaderboard at 1153, ahead of Ideogram v2 (1108) and Midjourney v6.1 (1100) [6][31].

At launch, BFL said fast high-resolution generation of up to 2K images was "coming soon" to the API. Launch API prices were 2.5 cents per image for FLUX.1 [dev], 5 cents for FLUX.1 [pro] and 4 cents for FLUX1.1 [pro], and BFL announced availability through Together.ai, Replicate, fal.ai and Freepik [6].

### FLUX1.1 [pro] Ultra and Raw Modes

On November 6, 2024, BFL released FLUX1.1 [pro] Ultra and Raw modes [16].

**Ultra mode** generates images at four times the resolution of standard FLUX1.1 [pro], up to 4 megapixels, in about 10 seconds per sample. BFL said its benchmarks showed Ultra was over 2.5 times faster than comparable high-resolution offerings. It was priced at $0.06 per image through the BFL API [16].

**Raw mode** is a toggle that, in BFL's words, "captures the genuine feel of candid photography", producing images with a less synthetic, more natural aesthetic. BFL said it significantly increases diversity in human subjects and enhances the realism of nature photography [16].

## FLUX.1 Tools

On November 21, 2024, BFL released FLUX.1 Tools, a suite of models that add control and steerability to the base FLUX.1 text-to-image model [17][55].

| Tool | Function | Availability at launch | Status (October 2026) |
|---|---|---|---|
| FLUX.1 Fill | Inpainting and outpainting from a text description and a binary mask | Pro (API) + Dev (open weights) | Active |
| FLUX.1 Depth | Structural guidance from a depth map extracted from an input image | Pro (API) + Dev (open weights, full and LoRA) | Deprecated |
| FLUX.1 Canny | Structural guidance from canny edges extracted from an input image | Pro (API) + Dev (open weights, full and LoRA) | Deprecated |
| FLUX.1 Redux | Adapter for mixing and recreating input images with text prompts | Pro (API, with FLUX1.1 [pro] Ultra) + Dev (open weights) | Active |

Each tool was released in the BFL API as a FLUX.1 [pro] variant and as a guidance-distilled, open-access FLUX.1 [dev] variant with inference code and weights, and was also offered through fal.ai, Replicate, Together.ai, Freepik and krea.ai. BFL reported that FLUX.1 Fill [pro] outperformed all competing methods on its inpainting benchmark, followed by FLUX.1 Fill [dev], and that FLUX.1 Depth outperformed Midjourney ReTexture [17]. BFL has since deprecated FLUX.1 Depth and FLUX.1 Canny ([pro], [dev] and LoRA versions) in its API; the open-weight checkpoints remain on Hugging Face for reference, while Fill and Redux are unaffected [17].

## NVIDIA collaboration

On January 2, 2025, BFL announced a collaboration with NVIDIA to optimize FLUX models for GeForce RTX 50 Series GPUs using FP4 compute on the Blackwell architecture. BFL said FLUX.1 [dev] then required only 10 GB of VRAM while running twice as fast on a GeForce RTX 5090 as on an RTX 4090 using plain BF16. The companies also announced a FLUX NIM microservice and an NVIDIA AI Blueprint for 3D-guided generative AI, in which a scene laid out in Blender guides FLUX image generation [39].

## FLUX Pro Finetuning API

On January 16, 2025, BFL launched the FLUX Pro Finetuning API, enabling users to customize FLUX.1 [pro] with their own images and concepts from as few as 1 to 5 example images. In a BFL user study, FLUX Pro finetuning results were preferred 68.9% of the time over other available finetuning services using FLUX.1 [dev] [18].

A finetune could be applied across FLUX.1 [pro], FLUX1.1 [pro] and the FLUX Tools suite without additional adaptation, including resolutions up to 4 megapixels, personalized inpainting with FLUX.1 Fill [pro] and structural control with FLUX.1 Depth [pro]. BFL announced German media group Burda (BurdaVerlag) as a launch partner [18]. The Finetuning API and all its endpoints were deprecated on October 31, 2025, with no migration path; in April 2026 BFL opened a separate public beta for serving user-trained LoRAs on FLUX.2 [klein] [43].

## FLUX.1 Kontext

On May 29, 2025, Black Forest Labs released FLUX.1 Kontext, a suite of flow matching models for in-context image generation and editing. Unlike standard [text-to-image models](https://aiwiki.ai/wiki/text-to-image_models), Kontext accepts both text and image inputs, allowing users to modify reference images through natural language instructions [37][12].

BFL listed the core capabilities as character consistency across scenes, local editing, style reference and interactive speed, and said users can build on previous edits over multiple turns. It also documented failure cases: long multi-turn editing sessions can introduce visual artifacts, the model occasionally ignores instructions, world knowledge is limited, and distillation can introduce artifacts [37]. The accompanying technical report describes a simple sequence concatenation approach that handles local editing and generative in-context tasks in one architecture, and introduces KontextBench, a benchmark of 1,026 image-prompt pairs across local editing, global editing, character reference, style reference and text editing [38].

| Model | Description | Availability |
|---|---|---|
| FLUX.1 Kontext [max] | Experimental model with improved prompt adherence, typography and editing consistency | API ($0.08/image) |
| FLUX.1 Kontext [pro] | Local editing, in-context modification and text-to-image in one model, built for fast iterative editing | API ($0.04/image) |
| FLUX.1 Kontext [dev] | Lightweight 12B diffusion transformer, editing only | [Open weights](https://aiwiki.ai/wiki/open_weights) ([Hugging Face](https://aiwiki.ai/wiki/hugging_face)), non-commercial license |

Prices are current BFL API list prices as of October 19, 2026 [25]. BFL reported that Kontext models run at inference speeds up to 8 times faster than current leading models such as GPT-Image [37]. At launch, Kontext [pro] and [max] were offered through KreaAI, Freepik, Lightricks, OpenArt, LeonardoAI and infrastructure partners including FAL, Replicate, Runware, DataCrunch, TogetherAI and ComfyOrg. BFL launched its web Playground at the same time [37].

On June 26, 2025, BFL released FLUX.1 Kontext [dev] as open weights under the FLUX.1 Non-Commercial License. BFL said it outperformed open editing models such as ByteDance Bagel and HiDream-E1-Full, and Google's closed Gemini-Flash Image, in many categories of KontextBench human preference evaluations, and released BF16, FP8 and FP4 TensorRT variants built with NVIDIA for Blackwell GPUs [36]. The same release introduced the self-serve commercial licensing portal and updated the FLUX.1 [dev] Non-Commercial License to require content filters or manual review and to require users to follow applicable content provenance law [36].

On September 25, 2025, FLUX.1 Kontext [pro] became available inside Generative Fill for Photoshop (beta) users, free for a limited time during the beta [41].

## FLUX.1 Krea [dev]

On July 31, 2025, BFL released FLUX.1 Krea [dev], developed in collaboration with [Krea AI](https://aiwiki.ai/wiki/krea_ai). It is a 12-billion-parameter rectified flow transformer and the open-weights version of Krea 1, trained to produce more realistic and diverse images without the oversaturated textures BFL calls the "AI look" [20][71]. BFL described it as "opinionated" and said it outperforms previous open text-to-image models and is "on par with closed solutions like FLUX1.1 [pro] in human preference assessments" [20].

It is architecturally compatible with FLUX.1 [dev], so it works as a drop-in replacement in systems built for that model. Weights are released under the FLUX.1 [dev] non-commercial license, with commercial licenses available through BFL's licensing portal [20][71].

## FLUX on Azure AI Foundry

On August 5, 2025, FLUX.1 Kontext [pro] and FLUX1.1 [pro] became available directly from [Microsoft](https://aiwiki.ai/wiki/microsoft) on Azure AI Foundry. BFL said its collaboration with Microsoft dated from its earliest days, when it partnered with [Azure](https://aiwiki.ai/wiki/azure) to build its training and inference clusters [40].

## Architecture

### How does the FLUX.1 architecture work?

BFL describes all public FLUX.1 models as based on "a hybrid architecture of multimodal and parallel [diffusion transformer](https://aiwiki.ai/wiki/diffusion_transformer) blocks" scaled to 12 billion parameters, building on flow matching and incorporating rotary positional embeddings and parallel attention layers [33]. BFL said at launch that it would publish a more detailed technical report. An unofficial reverse-engineering report by Or Greenberg (July 2025) states that the authors had not released official technical documentation of the architecture or training setup and documents the architecture from the public source code, and MLCommons notes that its architectural details and training procedures follow the Stable Diffusion 3 paper, "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis" [14][29][68]. MLCommons gives the model size as 11.9 billion parameters [29].

According to the configuration in BFL's reference code, the FLUX.1 transformer has 57 blocks with a hidden size of 3,072 and 24 attention heads, divided into two types [1]:

- **19 double-stream blocks:** Text and image tokens keep separate weights (their own normalization, projections and MLPs), but their queries, keys and values are concatenated and processed by a single joint attention operation, so information flows in both directions between the modalities [1]. This follows the MM-DiT design introduced in the Stable Diffusion 3 paper, which "uses separate weights for the two modalities and enables a bidirectional flow of information between image and text tokens" [68].
- **38 single-stream blocks:** Text and image tokens are processed as one concatenated sequence with shared weights. The reference code describes these as DiT blocks "with parallel linear layers", citing the parallel attention and MLP formulation from ViT-22B (arXiv 2302.05442) [1].

This hybrid design keeps modality-specific weights in the early layers and moves to fully shared processing in the later layers.

### Dual Text Encoders (FLUX.1)

FLUX.1 uses two text encoders, as configured in the reference code [1]:

- **[CLIP](https://aiwiki.ai/wiki/clip) ViT-L/14 (openai/clip-vit-large-patch14):** loaded with a maximum length of 77 tokens [1].
- **[T5](https://aiwiki.ai/wiki/t5) v1.1 XXL (google/t5-v1_1-xxl):** loaded with a default maximum length of 512 tokens; the FLUX.1 [schnell] model card's example uses a maximum sequence length of 256 [1][34].

### What is flow matching?

Flux uses flow matching as its training paradigm. BFL describes flow matching as "a general and conceptually simple method for training generative models, which includes diffusion as a special case" [33]. Flow matching, introduced by Lipman et al. (2022), trains a network to regress the velocity field of a continuous path that transports samples from a simple noise distribution to the data distribution [69]. Rectified flow, introduced by Liu, Gong and Liu (2022), is a formulation in which this path is a straight line between noise and data [70][68].

In traditional [diffusion model](https://aiwiki.ai/wiki/diffusion_model)s, training corrupts images with Gaussian noise and a network learns to reverse that process step by step. The Stable Diffusion 3 paper argues that rectified flow has better theoretical properties and conceptual simplicity, and reports that its improved rectified flow training outperformed established diffusion formulations for high-resolution text-to-image synthesis in a large-scale study [68]. MLCommons summarizes the practical difference for FLUX.1 as a shift to rectified flows, "which encourage linear denoising trajectories for more efficient decoding" [29].

### Rotary Positional Embeddings

FLUX incorporates [rotary positional embeddings](https://aiwiki.ai/wiki/rotary_position_embedding) (RoPE), which BFL says increase model performance and improve hardware efficiency [33]. In the reference code, the positional embedding is split across three axes with dimensions 16, 56 and 56 [1].

### Parallel Attention Layers

The single-stream blocks use parallel attention: rather than computing [self-attention](https://aiwiki.ai/wiki/self_attention) and the feedforward MLP sequentially, they are computed from the same input in parallel and their outputs combined [1][33]. BFL cites the technique as a hardware-efficiency improvement [33].

### Latent Space

Like Stable Diffusion, Flux operates in a compressed latent space. Images are encoded by a [variational autoencoder](https://aiwiki.ai/wiki/variational_autoencoder) (VAE), the flow model generates a latent, and the VAE decoder maps it back to pixels [66][14]. FLUX.1's autoencoder uses 16 latent channels in the reference configuration [1]. In the Stable Diffusion 3 study, 16-channel autoencoders gave clearly better reconstruction than 4- and 8-channel versions (for example an FID of 1.06 versus 2.41 for 4 channels), and the authors chose 16 channels for their models [68].

### FLUX.2: Next-Generation Architecture

FLUX.2 builds on a latent flow matching architecture and combines image generation and editing in a single architecture. It couples the Mistral-3 24B parameter vision-language model (VLM) with a rectified flow transformer; BFL says the VLM brings real-world knowledge and contextual understanding, while the transformer captures spatial relationships, material properties and composition [8]. The open-weight FLUX.2 [dev] is a 32-billion-parameter rectified flow transformer [46]. Black Forest Labs reports that FLUX.2 [dev] "sets a new standard, achieving leading performance across text-to-image generation, single-reference editing, and multi-reference editing, consistently outperforming all open-weights alternatives by a significant margin" [8].

BFL re-trained the model's latent space from scratch, describing the new FLUX.2 VAE as an optimized trade-off between learnability, quality and compression rate. The VAE was released on Hugging Face under the Apache 2.0 license [8].

For the FLUX.2 [klein] models, BFL replaced Mistral-3 with smaller Qwen3 text embedders. The 9B model uses a 9B flow model with an 8B Qwen3 text embedder, step-distilled to 4 inference steps [22][44]; the 4B model's published text encoder configuration is a smaller Qwen3 model [45].

### FLUX 3: Multimodal Architecture

BFL describes FLUX 3 as one multimodal foundation model that jointly learns from images, videos and audio, building on its Self-Flow approach for aligning multimodal generation and understanding in one architecture [48]. BFL wrote that video prediction accounts for over 95% of FLUX 3's total training compute, while audio makes up less than 0.5% of the tokens in a 720p video with audio [49]. The same backbone is used for content creation (FLUX 3 Video and FLUX 3 Image) and for action prediction in robotics (FLUX-mimic and FLUX 3 Action) [48][49]. FLUX 3 Image Klein uses a 5.6B image transformer with Qwen3-VL-4B-Instruct as its text encoder [53].

## FLUX.2 Models

On November 25, 2025, Black Forest Labs announced the FLUX.2 series. The announcement made FLUX.2 [pro], [flex] and [dev] available immediately, along with the FLUX.2 VAE, and listed FLUX.2 [klein] as "coming soon" [8]. All FLUX.2 variants offer image editing from text and multiple references in one model [8].

### Model Lineup

| Model | Parameters | Text Encoder | License | Release Date | Key Features |
|---|---|---|---|---|---|
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [max] | Not disclosed | Not disclosed | Proprietary (API) | December 16, 2025 | Highest quality, grounding search with real-time web information |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [pro] | Not disclosed | Not disclosed | Proprietary (API) | November 25, 2025 | Production-grade, multi-reference editing |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [flex] | Not disclosed | Not disclosed | Proprietary (API) | November 25, 2025 | Adjustable steps and guidance, typography specialist |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [dev] | 32B | Mistral-3 24B VLM | FLUX Non-Commercial License | November 25, 2025 | [Open weights](https://aiwiki.ai/wiki/open_weights), LoRA training, local deployment |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [klein] 9B (and Base 9B) | 9B | Qwen3 8B | FLUX Non-Commercial License | January 15, 2026 | Sub-second generation, 4-step distilled (Base is undistilled) |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [klein] 4B (and Base 4B) | 4B | Qwen3 | Apache 2.0 | January 15, 2026 | Smallest model, ~13 GB VRAM, consumer GPUs |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [klein] 9B-KV | 9B | Qwen3 8B | FLUX Non-Commercial License | March 2026 | KV-cache for up to 2.5x faster multi-reference editing |

Sources: [8][22][24][43][44][45][46].

### FLUX.2 [max]

FLUX.2 [max] launched on December 16, 2025 as BFL's highest-quality FLUX.2 model. BFL says it offers the highest editing consistency across tasks, the strongest prompt following and faithful style representation, and supports grounding search, generating images grounded in real-time information from the web such as current weather or recent news [23][43]. It combines up to 10 input images while maintaining identity [43].

### FLUX.2 [pro]

FLUX.2 [pro] is the production-grade variant. At launch, BFL priced it from $0.03 per megapixel and said the API combines up to 8 input images (10 in the Playground) [43]; the FLUX.2 announcement describes up to 10 reference images with "the best character / product / style consistency available today" [8]. Output goes up to 4 megapixels [8][43]. On March 3, 2026, BFL made FLUX.2 [pro] about twice as fast with no price change, and added a flux-2-pro-preview endpoint where new improvements land first while the original endpoint stays a fixed snapshot [43].

### FLUX.2 [flex]

FLUX.2 [flex] exposes inference parameters such as the number of steps (up to 50) and the guidance scale (1.5 to 10), letting developers trade off quality, prompt adherence and speed. BFL describes it as excelling at text rendering and fine details [8][43]. On January 29, 2026, BFL said [flex] had become up to 3 times faster with the same quality [43].

### FLUX.2 [dev]

The open-weight 32B model combines text-to-image synthesis and image editing with multiple input images in a single checkpoint. Weights are on Hugging Face, and BFL released an optimized fp8 reference implementation for GeForce RTX GPUs, created with NVIDIA and [ComfyUI](https://aiwiki.ai/wiki/comfyui). API endpoints for [dev] were offered by FAL, Replicate, Runware, Verda, TogetherAI, Cloudflare and DeepInfra [8]. BFL says the open-weight release decision followed a third-party evaluation for synthetic CSAM and NCII of about 2,800 prompts [46].

### FLUX.2 [klein]

Klein (German for "small") is the FLUX.2 family's fastest tier, released on January 15, 2026. BFL says it unifies text-to-image, image editing and multi-reference generation in one compact model and can generate or edit images in under 0.5 seconds on modern hardware. The 4B model fits in about 13 GB of VRAM, on GPUs such as the RTX 3090 or 4070 and above [22].

The licenses differ by size. FLUX.2 [klein] 4B and Base 4B are released under Apache 2.0, while FLUX.2 [klein] 9B and Base 9B are open-weight under the FLUX Non-Commercial License [22][43][44][45]. With this release, BFL renamed the "FLUX [dev] Non-Commercial License" to the "FLUX Non-Commercial License", saying no material changes had been made [22]. The Base variants are undistilled and intended for fine-tuning, LoRA training and research. BFL also released FP8 and NVFP4 versions of all [klein] variants, developed with NVIDIA, reporting up to 1.6x speed and 40% less VRAM for FP8 and up to 2.7x speed and 55% less VRAM for NVFP4 on RTX 5080/5090 [22].

In March 2026, BFL published FLUX.2 [klein] 9B-KV and its FP8 variant on Hugging Face. By caching key-value pairs for reference image tokens during the first denoising step and reusing them in later steps, it speeds up multi-reference editing by up to 2.5 times; it is also released under the FLUX Non-Commercial License [24].

## FLUX 3

### Announcement and rollout

On July 23, 2026, BFL announced FLUX 3 in early access, describing it as a multimodal foundation model and "a checkpoint on our mission to develop real-world visual intelligence: models that perceive, predict, and act across physical and digital environments" [48]. BFL's launch plan listed four products, each to follow an early access phase: FLUX 3 Video (video and audio generation and editing through APIs and private weight access), FLUX-mimic and FLUX 3 Action (action prediction with selected partners, beginning with [mimic robotics](https://aiwiki.ai/wiki/mimic_robotics)), FLUX 3 Image (image synthesis and editing), and FLUX 3 Dev (open-weight access to a multimodal backbone) [48].

In preliminary BFL evaluations of 10-second 720p text-to-video clips with audio, FLUX 3 was preferred over Luma Ray 3.2 in 93% of comparisons, Runway Gen-4.5 in 77%, Grok Imagine Video in up to 69%, Kling v3 Pro in 60%, and Seedance 2.0 and Gemini Omni Flash in 52% [48]. The Decoder noted that BFL called the results preliminary and that no independent tests were available at the time [54].

| Product | Date | Notes |
|---|---|---|
| FLUX 3 (early access) | July 23, 2026 | Multimodal model; video with native audio up to 20 seconds [48] |
| FLUX-mimic | July 23, 2026 | Video-action model with mimic robotics, tested and deployed on robots at Audi [49] |
| [FLUX 3 Video](https://aiwiki.ai/wiki/flux_3_video) | August 4, 2026 | Generally available via API; clips up to 20 seconds at HD with Full HD via upscaling and native audio [50] |
| FLUX 3 Action | September 2026 | Open-weight 7B world action model for robot control [51] |
| [FLUX 3 Image](https://aiwiki.ai/wiki/flux_3_image) | October 1, 2026 | API; bounding boxes in the prompt, up to ten references, up to 4K output [43] |
| FLUX 3 Image Klein | October 15, 2026 | 5.6B open-weight image model, base and distilled checkpoints [52] |

### FLUX 3 Video

FLUX 3 Video became generally available through the BFL API and selected partners on August 4, 2026, generating clips of up to 20 seconds with synchronized audio, including multilingual dialogue with lip-sync [50][43]. It supports text-to-video, image-to-video with keyframes, video continuation and a low-cost draft mode, and BFL said its internal evaluations showed it tying Seedance 2.0 and beating other state-of-the-art models in image-to-video [50]. BFL later capped video continuation at 15 seconds and, on September 10, 2026, added 2K and 4K output [43]. See [FLUX 3 Video](https://aiwiki.ai/wiki/flux_3_video) for details.

### FLUX 3 Action

FLUX 3 Action is an open-weight 7B world action model derived from the FLUX 3 backbone. BFL reports that it sets a new state-of-the-art success rate on the RoboLab-120 simulation benchmark (38.3% for the single-step checkpoint and 42.2% for the guidance-distilled checkpoint, against 36.8% for NVIDIA's Cosmos 3 Nano) while running faster than comparable world action models [51]. Its Hugging Face repositories were published under the FLUX Kommunity License [72].

### FLUX 3 Image and FLUX 3 Image Klein

FLUX 3 Image was released in the BFL API on October 1, 2026. It handles generation and editing through one endpoint, accepts bounding boxes written into the prompt to place or confine elements, combines up to ten reference images, and outputs from 768 x 768 up to 4K (about 16 megapixels). Launch prices ranged from $0.041 per image at 768sq to $0.607 at 4K [43]. See [FLUX 3 Image](https://aiwiki.ai/wiki/flux_3_image) for details.

FLUX 3 Image Klein, released on October 15, 2026, is a 5.6-billion-parameter open-weight version supporting text-to-image, image editing, multi-reference generation with up to 10 references and bounding boxes. It ships as a step-distilled model (16 steps) and an undistilled Base model (50 steps) for fine-tuning, with FP8, INT8 and NVFP4 builds developed with NVIDIA [52][43]. BFL says the generator weights need 11.2 GB in BF16 and 5.6 GB in FP8, and that the INT8 build runs on GeForce RTX GPUs with as little as 8 GB of VRAM [52]. In BFL's internal evaluations, it ranked first among the compared models in text-to-image, multi-reference generation and image editing, and gained 91 Elo in text-to-image and 70 in image editing over FLUX.2 [klein] 9B with 38% fewer parameters [52]. API prices start at $0.008 per image at 768sq and $0.009 at 1K [43]. The model card says the model should be used with a prompt upsampler, and BFL publishes its own as a separate 9B model [53].

## Reception and benchmarks

At the FLUX.1 launch, TechCrunch reported that FLUX.1 models surpassed Midjourney's and OpenAI's image generators in quality as ranked by users in the Artificial Analysis image arena [2]. FLUX1.1 [pro] then led the Artificial Analysis leaderboard with an Elo of 1153 as of October 1, 2024 [31]. MLCommons, explaining its choice of FLUX.1 as an MLPerf benchmark in October 2025, wrote that it "can compete with closed-source models across multiple quality benchmarks" [29]. Most other performance claims for FLUX models, including the comparisons for FLUX.2, FLUX 3 Video and FLUX 3 Image Klein above, come from BFL's own evaluations.

Flux's main differentiator against closed competitors such as [Midjourney](https://aiwiki.ai/wiki/midjourney) and [DALL-E 3](https://aiwiki.ai/wiki/dall_e_3) is that several variants are released as open weights, enabling local deployment, fine-tuning and community extensions [33][8][22].

## API Access and Pricing

Black Forest Labs offers a credit-based API in which 1 credit equals $0.01 USD; the same price applies in the API and the Playground. FLUX.2 and FLUX 3 Image models are priced by output resolution, while older FLUX.1 models have flat per-image prices [25].

| Model | Price per image (as of October 19, 2026) | Notes |
|---|---|---|
| FLUX 3 Image Klein | From $0.008 (768sq) to $0.036 (4K) | Same price for generation and editing [43] |
| [FLUX 3 Image](https://aiwiki.ai/wiki/flux_3_image) | From $0.041 (768sq) to $0.607 (4K) | Launch pricing [43] |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [klein] 4B | From $0.014 | Megapixel-based pricing [25] |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [klein] 9B | From $0.015 | Megapixel-based pricing [25] |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [pro] | From $0.03 (generation), from $0.045 (editing) | Production-grade [25] |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [max] | From $0.07 (generation), from $0.10 (editing) | Highest FLUX.2 quality [25] |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [flex] | From $0.05 (generation), from $0.10 (editing) | Tunable parameters [25] |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) [dev] | No hosted BFL API | Open weights, non-commercial local use [25] |
| FLUX.1 Kontext [pro] | $0.04 | Context-aware editing [25] |
| FLUX.1 Kontext [max] | $0.08 | Highest Kontext quality [25] |
| FLUX1.1 [pro] | $0.04 | Standard generation [25] |
| FLUX1.1 [pro] Ultra | $0.06 | 4MP high-resolution [25] |
| FLUX.1 Fill [pro] | $0.05 | Inpainting/outpainting [25] |

FLUX 3 Video is priced per second of output [25][43]. Batch requests multiply the base cost by the number of images requested [25]. FLUX models are also available through third-party platforms, including [Together AI](https://aiwiki.ai/wiki/together_ai), [Replicate](https://aiwiki.ai/wiki/replicate), [fal.ai](https://aiwiki.ai/wiki/fal_ai), Runware, DataCrunch, [Cloudflare](https://aiwiki.ai/wiki/cloudflare), [DeepInfra](https://aiwiki.ai/wiki/deepinfra) and Azure AI Foundry, and through NVIDIA's FLUX NIM microservice [6][8][37][39][40].

## Adoption and Integrations

### Notable Integrations

- **Grok (xAI):** On August 13, 2024, xAI launched image generation in Grok on X powered by FLUX.1, with few safeguards against misuse at launch [2][58]. On December 7, 2024, The Verge reported that X had given Grok a new image generator called Aurora, in a "Grok 2 + Aurora beta" option alongside Grok's existing generator [59]. In September 2025, Trending Topics reported that a source close to BFL had said in April 2025 that the xAI partnership had ended [57].
- **Mistral Le Chat:** On November 18, 2024, Mistral AI added image generation powered by Black Forest Labs' Flux Pro to Le Chat [60].
- **Freepik:** The design platform integrated FLUX into its AI Suite, scaling to millions of inference requests per day with infrastructure partners DataCrunch and WaveSpeed; DataCrunch says Freepik's customers generate over 60 million images per month [11].
- **Burda:** German media group BurdaVerlag was a launch partner for the FLUX Pro Finetuning API in January 2025 [18].
- **Microsoft Azure AI Foundry:** On August 5, 2025, FLUX.1 Kontext [pro] and FLUX1.1 [pro] became available directly from Microsoft on Azure AI Foundry [40].
- **Meta:** Meta used BFL models in the early version of its Vibes AI video feed, launched in September 2025, and reportedly signed a multi-year contract worth $35 million in its first year and $105 million in its second [61][57].
- **Adobe Photoshop:** From September 25, 2025, FLUX.1 Kontext [pro] became available as a model in Photoshop (beta) Generative Fill [41].
- **Film:** In June 2026, director Martin Scorsese joined BFL as an adviser and was shown using FLUX to storyboard a scene; Variety reported that he was introduced to the company through BroadLight Capital, a BFL investor co-founded by his manager Rick Yorn [62].

### Community and ComfyUI

The open-weight FLUX models have a large community around [ComfyUI](https://aiwiki.ai/wiki/comfyui), the node-based workflow tool, which had day-one integration for FLUX.1 [33]. BFL has also shipped day-zero support in ComfyUI, Hugging Face Diffusers and TensorRT for FLUX.1 Kontext [dev], and day-one support in ComfyUI, Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni for FLUX 3 Image Klein [36][52]. In November 2025, NVIDIA said it had worked with BFL and ComfyUI to ship FLUX.2 with FP8 quantizations that reduce VRAM requirements by 40% and improve performance by 40% on RTX GPUs [26].

FLUX models are widely fine-tuned with [LoRA](https://aiwiki.ai/wiki/lora) adapters. BFL's own FLUX.2 [klein] Base and FLUX 3 Image Klein Base checkpoints are released specifically for fine-tuning and LoRA training, and BFL lists Ostris's AI Toolkit and NVIDIA NeMo AutoModel recipes for FLUX 3 Image Klein [22][52].

### Fal.AI FLUX.2 [dev] Turbo

On December 29, 2025, Fal.AI released FLUX.2 [dev] Turbo on Hugging Face, a distilled LoRA adapter for FLUX.2 [dev] that enables image generation in 8 inference steps, which Fal describes as 6x faster than the base model's typical 50 steps [65].

## MLPerf Benchmark

In October 2025, MLCommons announced that MLPerf Training v5.1 would use Flux.1 as its new text-to-image benchmark, replacing Stable Diffusion v2 to reflect modern model architectures and scale. MLCommons describes the benchmarked model as an 11.9-billion-parameter transformer-based model. Because the original Flux.1 training code is not open source, the reference implementation builds on the torchtitan implementation, and the benchmark trains from scratch on the CC12M dataset until a target validation loss [29]. In the MLPerf Training v5.1 results published in November 2025, NVIDIA set a record time-to-train of 12.5 minutes using 1,152 Blackwell GPUs and was the only platform to submit to the benchmark [32].

## Safety and controversies

The Grok integration in August 2024 drew the first wave of criticism. TechCrunch reported that Grok's FLUX-powered image generator had very few safeguards, allowing users to post fake images of public figures such as Donald Trump directly to X, and that images like Pikachu holding an assault rifle could not be re-created with Google's or OpenAI's generators [2]. The Verge described the feature as having "seemingly few guardrails in place to prevent abuse" [58].

In June 2026, the Art Directors Guild criticized Martin Scorsese's work with BFL, arguing that generative AI is built by "ingesting large swaths of copyrighted work, likely scraped from the internet without consent, credit, compensation, or transparency" [63].

BFL has published its own mitigation approach. In a February 2026 post, it said it filters pre-training data for nude and pornographic material and known CSAM (partnering with the Internet Watch Foundation), applies targeted post-training fine-tuning against synthetic NCII and CSAM, requires filters in its open-weight licenses, applies C2PA content provenance metadata on hosted services, and had Cinder red-team each of its five FLUX.2 open-weight releases. BFL reported that its post-training mitigations cut vulnerabilities by 77-98% and that its models showed over 10 times fewer vulnerabilities than other popular open-weight models, including models from Alibaba, Tencent and ByteDance; these are BFL's reported results from evaluations it commissioned [47]. BFL said FLUX 3 Video and FLUX 3 Image Klein were also evaluated by Cinder before release [50][52]. In a speech at the G7 that BFL published on June 18, 2026, CEO Robin Rombach described BFL as "a 90-person frontier AI lab" and urged governments to support open and responsible AI development [64].

## Is Flux open source?

Black Forest Labs uses a tiered licensing strategy, so the answer depends on the variant: some models are fully open source under Apache 2.0, several are open-weight under non-commercial or community licenses, and the top models are proprietary and API-only [33][22][52].

| License | Variants |
|---|---|
| Apache 2.0 | FLUX.1 [schnell]; FLUX.2 [klein] 4B and Base 4B; FLUX.2 VAE [34][22][8] |
| FLUX.1 [dev] Non-Commercial License | FLUX.1 [dev], FLUX.1 Tools [dev], FLUX.1 Kontext [dev], FLUX.1 Krea [dev] [35][17][36][71] |
| FLUX Non-Commercial License (formerly FLUX [dev] Non-Commercial License) | FLUX.2 [dev], FLUX.2 [klein] 9B, Base 9B and 9B-KV [46][22][24] |
| FLUX Kommunity License v1.0 | FLUX 3 Image Klein and Base; FLUX 3 Action [53][72] |
| Proprietary (API only) | FLUX.1 [pro], FLUX1.1 [pro] and Ultra, FLUX.1 Kontext [pro] and [max], FLUX.2 [pro], [flex] and [max], FLUX 3 Video and FLUX 3 Image [33][37][8][43] |

- **Apache 2.0** allows commercial and personal use, modification and redistribution [34][45].
- **FLUX.1 [dev] Non-Commercial License and FLUX Non-Commercial License** allow free use of the weights for research and non-commercial purposes; commercial use of the models requires a separate license, which BFL sells through its self-serve portal. The FLUX.1 [dev] model card states that generated outputs may be used for commercial purposes as described in that license, and since June 2025 the FLUX.1 [dev] license requires content filters or manual review [35][36].
- **FLUX Kommunity License** permits non-commercial use plus commercial output generation by "Qualifying Users" with less than US$5 million in gross annualized revenue, including affiliates; larger organizations use BFL's self-hosted licensing or the API [53][52].

## Technical Lineage

Flux represents the technical evolution of ideas from several research papers:

1. **High-Resolution Image Synthesis with [Latent Diffusion Model](https://aiwiki.ai/wiki/latent_diffusion)s** (Rombach, Blattmann, Lorenz, Esser and Ommer, 2021/2022): established diffusion in a compressed autoencoder latent space, the basis for Stable Diffusion [66].
2. **Scalable Diffusion Models with Transformers** (Peebles and Xie, 2022 preprint): introduced the [Diffusion Transformer (DiT)](https://aiwiki.ai/wiki/diffusion_transformer) architecture, replacing the [U-Net](https://aiwiki.ai/wiki/unet) backbone with transformer blocks [67].
3. **Flow Matching for Generative Modeling** (Lipman et al., 2022) and **Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow** (Liu, Gong and Liu, 2022): the training frameworks behind FLUX's flow-based formulation [69][70].
4. **Scaling Rectified Flow Transformers for High-Resolution Image Synthesis** (Esser et al., 2024): the Stable Diffusion 3 paper describing the MM-DiT architecture and rectified flow training, which MLCommons says FLUX.1 follows; first-authored by BFL co-founder Patrick Esser [68][29].
5. **FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space** (Black Forest Labs, 2025): BFL's technical report on unified generation and editing, introducing KontextBench [38].

## Release Timeline

| Date | Release |
|---|---|
| August 1, 2024 | [Black Forest Labs](https://aiwiki.ai/wiki/black_forest_labs) launch; [FLUX.1](https://aiwiki.ai/wiki/flux_1) [schnell], [dev], [pro] released |
| August 13, 2024 | FLUX.1 powers image generation in xAI's Grok |
| October 2, 2024 | FLUX1.1 [pro] and BFL API general availability (beta) |
| November 6, 2024 | FLUX1.1 [pro] Ultra and Raw modes |
| November 18, 2024 | Flux Pro image generation in Mistral's Le Chat |
| November 21, 2024 | FLUX.1 Tools (Fill, Depth, Canny, Redux) |
| January 2, 2025 | NVIDIA collaboration for RTX 50 Series and FP4 optimization |
| January 16, 2025 | FLUX Pro Finetuning API launch |
| May 29, 2025 | FLUX.1 Kontext ([pro], [max]) and BFL Playground |
| June 26, 2025 | FLUX.1 Kontext [dev] open weights; self-serve licensing portal |
| July 31, 2025 | FLUX.1 Krea [dev] (collaboration with [Krea AI](https://aiwiki.ai/wiki/krea_ai)) |
| August 5, 2025 | FLUX.1 Kontext [pro] and FLUX1.1 [pro] on Azure AI Foundry |
| September 25, 2025 | FLUX.1 Kontext [pro] in Adobe Photoshop (beta) Generative Fill |
| October 31, 2025 | FLUX.1 [pro] 1.0 endpoints and Finetuning API deprecated |
| November 25, 2025 | [FLUX.2](https://aiwiki.ai/wiki/flux_2) [pro], [flex], [dev] and FLUX.2 VAE |
| December 1, 2025 | $300M Series B at $3.25B post-money valuation |
| December 16, 2025 | [FLUX.2](https://aiwiki.ai/wiki/flux_2) [max] |
| January 15, 2026 | [FLUX.2](https://aiwiki.ai/wiki/flux_2) [klein] 4B and 9B (plus Base variants) |
| February 24, 2026 | BFL publishes FLUX.2 safety evaluation results |
| March 2026 | [FLUX.2](https://aiwiki.ai/wiki/flux_2) [klein] 9B-KV and FP8 variant |
| June 2026 | Martin Scorsese joins BFL as adviser |
| July 23, 2026 | FLUX 3 announced (early access) and FLUX-mimic |
| August 4, 2026 | [FLUX 3 Video](https://aiwiki.ai/wiki/flux_3_video) generally available |
| September 2026 | FLUX 3 Action (open weights, 7B) |
| October 1, 2026 | [FLUX 3 Image](https://aiwiki.ai/wiki/flux_3_image) in the BFL API |
| October 15, 2026 | FLUX 3 Image Klein (open weights, 5.6B) |

Sources: [33][58][6][16][60][17][39][18][37][36][20][40][41][43][8][42][22][47][24][62][48][49][50][51][52].

## Current State (October 2026)

As of October 19, 2026, the FLUX family spans three generations. BFL's current hosted lineup includes FLUX 3 Video and FLUX 3 Image at the top end, the FLUX.2 [pro], [max], [flex] and [klein] image models, and legacy FLUX.1 Kontext and FLUX1.1 [pro] endpoints, alongside newer FLUX Tools endpoints for outpainting, object removal, deblurring, virtual try-on and video upscaling [25][43]. Open-weight releases range from the Apache 2.0 FLUX.1 [schnell] and FLUX.2 [klein] 4B to the community-licensed FLUX 3 Image Klein and FLUX 3 Action [22][51][52]. BFL's July 2026 launch plan also lists an open-weight FLUX 3 Dev multimodal backbone that had not yet been released as of the FLUX 3 Image Klein announcement, which described Klein as the second open-weight FLUX 3 release after FLUX 3 Action [48][52].

The company has broadened its stated focus from image generation to what it calls "visual intelligence", covering video, audio and action prediction for robotics [48][49].

## See Also

- [Black Forest Labs](https://aiwiki.ai/wiki/black_forest_labs)
- [FLUX.1](https://aiwiki.ai/wiki/flux_1)
- [FLUX.2](https://aiwiki.ai/wiki/flux_2)
- [FLUX 3 Image](https://aiwiki.ai/wiki/flux_3_image)
- [FLUX 3 Video](https://aiwiki.ai/wiki/flux_3_video)
- [Stable Diffusion](https://aiwiki.ai/wiki/stable_diffusion)
- [Rectified Flow](https://aiwiki.ai/wiki/rectified_flow)
- [Diffusion Transformer](https://aiwiki.ai/wiki/diffusion_transformer)
- [AI image generation](https://aiwiki.ai/wiki/ai_image_generation)
- [Generative AI](https://aiwiki.ai/wiki/generative_ai)
- [Transformer](https://aiwiki.ai/wiki/transformer)
- [Latent diffusion](https://aiwiki.ai/wiki/latent_diffusion)
- [Midjourney](https://aiwiki.ai/wiki/midjourney)
- [DALL-E](https://aiwiki.ai/wiki/dall-e)
- [ComfyUI](https://aiwiki.ai/wiki/comfyui)

## References

1. [Official inference repo for FLUX.1 models](https://github.com/black-forest-labs/flux) - GitHub, Black Forest Labs
2. [Meet Black Forest Labs, the startup powering Elon Musk's unhinged AI image generator](https://techcrunch.com/2024/08/14/meet-black-forest-labs-the-startup-powering-elon-musks-unhinged-ai-image-generator/) - TechCrunch, August 2024
3. [Black Forest Labs raises $300M at $3.25B valuation](https://techcrunch.com/2025/12/01/black-forest-labs-raises-300m-at-3-25b-valuation/) - TechCrunch, December 2025
4. [Stable Diffusion creators launch Black Forest Labs, secure $31M for FLUX.1 AI image generator](https://venturebeat.com/ai/stable-diffusion-creators-launch-black-forest-labs-secure-31m-for-flux-1-ai-image-generator) - VentureBeat, August 2024
5. [Exclusive: Black Forest Labs is raising another $100M on a $1B valuation](https://techcrunch.com/2024/09/20/grok-image-generator-black-forest-labs-raising-100m-at-1b-valuation/) - TechCrunch, September 2024
6. [Announcing FLUX1.1 [pro] and the BFL API](https://bfl.ai/announcing-flux-1-1-pro-and-the-bfl-api/) - Black Forest Labs, October 2024
7. [How does Flux work? The new image generation AI that rivals Midjourney](https://medium.com/@drmarcosv/how-does-flux-work-the-new-image-generation-ai-that-rivals-midjourney-7f81f6f354da) - Marcos V. Conde, Medium
8. [FLUX.2: Frontier Visual Intelligence](https://bfl.ai/blog/flux-2) - Black Forest Labs, November 2025
9. [FLUX.1 vs Midjourney: Text to Image AI Showdown](https://getimg.ai/blog/flux-1-vs-midjourney-ultimate-text-to-image-ai-showdown) - getimg.ai
10. [FLUX.1 AI Image Gen in Grok 2.0](https://www.ultralytics.com/blog/xai-launches-grok-2-0-with-flux-1-integration) - Ultralytics
11. [How Freepik scaled FLUX media generation to millions of requests per day](https://datacrunch.io/blog/how-freepik-scaled-flux-media-generation-to-millions-of-requests-per-day) - DataCrunch
12. [Black Forest Labs' Kontext AI models can edit pics as well as generate them](https://techcrunch.com/2025/05/29/black-forest-labs-kontext-ai-models-can-edit-pics-as-well-as-generate-them/) - TechCrunch, May 2025
13. [Black Forest Labs launches open source Flux.2 [klein] to generate AI images in less than a second](https://venturebeat.com/technology/black-forest-labs-launches-open-source-flux-2-klein-to-generate-ai-images-in) - VentureBeat, January 2026
14. [Demystifying Flux Architecture](https://arxiv.org/html/2507.09595v1) - arXiv
15. [Meta to Pay $140 Million to Use Black Forest Labs AI for Images](https://sifted.eu/articles/black-forest-labs-meta) - Sifted / Bloomberg, September 2025
16. [Introducing FLUX1.1 [pro] Ultra and Raw Modes](https://bfl.ai/flux-1-1-ultra/) - Black Forest Labs, November 2024
17. [Introducing FLUX.1 Tools](https://bfl.ai/flux-1-tools/) - Black Forest Labs, November 2024
18. [Announcing the FLUX Pro Finetuning API](https://bfl.ai/blog/25-01-16-finetuning) - Black Forest Labs, January 2025
19. [Black Forest Labs announcements](https://bfl.ai/announcements) - Black Forest Labs
20. [FLUX.1 Krea [dev]: An 'Opinionated' Text-to-Image Model](https://bfl.ai/blog/flux-1-krea-dev) - Black Forest Labs, July 2025
21. [FLUX.1-dev: Encoders and Token Limitations](https://medium.com/@lbq999/flux-1-dev-encoders-and-token-limitations-8631c179eaad) - Medium
22. [FLUX.2 [klein]: Towards Interactive Visual Intelligence](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence) - Black Forest Labs, January 2026
23. [FLUX.2 [max] - Top-Tier Quality Image Generation](https://bfl.ai/models/flux-2-max) - Black Forest Labs
24. [FLUX.2-klein-9b-kv](https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-kv) - Hugging Face, March 2026
25. [Pricing overview](https://docs.bfl.ml/quick_start/pricing) - Black Forest Labs Documentation
26. [FLUX.2 Image Generation Models Now Released, Optimized for NVIDIA RTX GPUs](https://blogs.nvidia.com/blog/rtx-ai-garage-flux-2-comfyui/) - NVIDIA Blog
27. [ComfyUI FLUX LoRA Training: Detailed Guides](https://www.runcomfy.com/comfyui-workflows/comfyui-flux-lora-training-detailed-guides) - RunComfy
28. [New Year's AI surprise: Fal releases its own version of Flux 2 image generator](https://venturebeat.com/technology/new-years-ai-surprise-fal-releases-its-own-version-of-flux-2-image-generator) - VentureBeat, December 2025
29. [MLPerf Training Introduces Flux.1 Text-to-Image Benchmark](https://mlcommons.org/2025/10/training-flux1/) - MLCommons, October 2025
30. [Into the new era of visual intelligence: Our investment in Black Forest Labs](https://northzone.com/2025/12/01/black-forest-labs/) - Northzone, December 2025
31. [Black Forest Labs releases Flux 1.1 Pro and an API](https://venturebeat.com/ai/black-forest-labs-releases-flux-1-1-pro-and-an-api) - VentureBeat, October 2024
32. [NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks](https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/) - NVIDIA Technical Blog, November 2025
33. [Announcing Black Forest Labs](https://bfl.ai/blog/24-08-01-bfl) - Black Forest Labs, August 1, 2024
34. [black-forest-labs/FLUX.1-schnell](https://huggingface.co/black-forest-labs/FLUX.1-schnell) - Hugging Face model card
35. [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) - Hugging Face model card
36. [FLUX.1 Kontext [dev] - Open Weights for Image Editing](https://bfl.ai/blog/flux-1-kontext-dev) - Black Forest Labs, June 26, 2025
37. [Introducing FLUX.1 Kontext and the BFL Playground](https://bfl.ai/blog/flux-1-kontext) - Black Forest Labs, May 29, 2025
38. [FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space](https://arxiv.org/abs/2506.15742) - Black Forest Labs et al., arXiv, June 2025
39. [Bringing Lightning-Fast FLUX Performance to More Creators in Collaboration with NVIDIA](https://bfl.ai/blog/25-01-03-nvidia) - Black Forest Labs, January 2, 2025
40. [FLUX Models Launch on Azure AI Foundry for Enterprise-Ready Image Generation](https://bfl.ai/blog/flux-azure-ai-foundry) - Black Forest Labs, August 5, 2025
41. [FLUX.1 Kontext now in Adobe Photoshop: Powering Every Pixel](https://bfl.ai/blog/flux1-kontext-now-in-adobe-photoshop-powering-every-pixel) - Black Forest Labs, September 25, 2025
42. [Laying the Foundations for Visual Intelligence: Our $300M Series B](https://bfl.ai/blog/our-300m-series-b) - Black Forest Labs, December 1, 2025
43. [Release Notes](https://docs.bfl.ai/release-notes) - Black Forest Labs Documentation, accessed October 19, 2026
44. [black-forest-labs/FLUX.2-klein-9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-9B) - Hugging Face model card, January 2026
45. [black-forest-labs/FLUX.2-klein-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B) - Hugging Face model card and repository, January 2026
46. [black-forest-labs/FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev) - Hugging Face model card, November 2025
47. [Capable, Open, and Safe: Combating AI Misuse](https://bfl.ai/blog/capable-open-and-safe-combating-ai-misuse) - Black Forest Labs, February 24, 2026
48. [FLUX 3 - Real World Models: Towards Multimodal Flow Models as the Backbone of Visual Intelligence](https://bfl.ai/blog/flux-3) - Black Forest Labs, July 23, 2026
49. [FLUX 3 x mimic: The Next Generation of Video-Action Models](https://bfl.ai/blog/flux-3-mimic) - Black Forest Labs, July 23, 2026
50. [FLUX 3 Video, Part 1: Generation](https://bfl.ai/blog/flux-3-video) - Black Forest Labs, August 4, 2026
51. [FLUX 3 Action: A 7B World Action Model for Robot Control](https://bfl.ai/models/flux-3-action) - Black Forest Labs, September 2026
52. [FLUX 3 Image Klein: open weights, 5.6B parameters](https://bfl.ai/blog/flux-3-image-klein) - Black Forest Labs, October 15, 2026
53. [black-forest-labs/flux-3-image-klein](https://huggingface.co/black-forest-labs/flux-3-image-klein) - Hugging Face model card, October 2026
54. [Flux 3 generates videos with native audio up to 20 seconds long, a first for Black Forest Labs](https://the-decoder.com/flux-3-generates-videos-with-native-audio-up-to-20-seconds-long-a-first-for-black-forest-labs/) - The Decoder, July 23, 2026
55. [Black Forest Labs expands FLUX.1 with four new AI tools for image editing](https://the-decoder.com/black-forest-labs-expands-flux-1-with-four-new-ai-tools-for-image-editing/) - The Decoder, November 22, 2024
56. [Germany's AI Image Generator Black Forest Labs Raises $300M At $3.25B Valuation As European AI Fundings Scale Up](https://news.crunchbase.com/ai/image-generator-europe-unicorn-black-forest-labs-raise/) - Crunchbase News, December 1, 2025
57. [Deutsches AI-Startup Black Forest Labs soll großen Deal mit Meta geschlossen haben](https://www.trendingtopics.eu/deutsche-ai-startup-black-forest-labs-soll-grosse-deal-mit-meta-geschlossen-haben/) - Trending Topics (citing Bloomberg), September 10, 2025
58. [xAI's new Grok-2 chatbots bring AI image generation to X](https://www.theverge.com/2024/8/14/24220127/grok-ai-chatbot-beta-image-generation-x-xai-update) - The Verge, August 14, 2024
59. [X gives Grok a new photorealistic AI image generator](https://www.theverge.com/2024/12/7/24315644/grok-x-aurora-ai-image-generator-xai) - The Verge, December 7, 2024
60. [Mistral has entered the chat](https://mistral.ai/news/mistral-chat/) - Mistral AI, November 18, 2024
61. [Meta launches 'Vibes,' a short-form video feed of AI slop](https://techcrunch.com/2025/09/25/meta-launches-vibes-a-short-form-video-feed-of-ai-slop/) - TechCrunch, September 25, 2025
62. [Martin Scorsese Backs AI Company and Says He's Using It to Storyboard Movies](https://variety.com/2026/film/news/martin-scorsese-supports-ai-company-storyboard-movies-1236765037/) - Variety, June 2, 2026
63. [Martin Scorsese's AI Support Draws Pushback From Art Directors Guild](https://www.yahoo.com/entertainment/movies/articles/martin-scorsese-ai-support-draws-144004061.html) - Yahoo Entertainment, June 2026
64. [Our co-founder and CEO urges G7 leaders to back open innovation](https://bfl.ai/blog/our-co-founder-and-ceo-urges-g7-leaders-to-back-open-innovation) - Black Forest Labs, June 18, 2026
65. [fal/FLUX.2-dev-Turbo](https://huggingface.co/fal/FLUX.2-dev-Turbo) - Hugging Face model card, December 2025
66. [High-Resolution Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) - Rombach, Blattmann, Lorenz, Esser, Ommer, arXiv, December 2021
67. [Scalable Diffusion Models with Transformers](https://arxiv.org/abs/2212.09748) - Peebles and Xie, arXiv, December 2022
68. [Scaling Rectified Flow Transformers for High-Resolution Image Synthesis](https://arxiv.org/abs/2403.03206) - Esser et al., arXiv, March 2024
69. [Flow Matching for Generative Modeling](https://arxiv.org/abs/2210.02747) - Lipman et al., arXiv, October 2022
70. [Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow](https://arxiv.org/abs/2209.03003) - Liu, Gong, Liu, arXiv, September 2022
71. [black-forest-labs/FLUX.1-Krea-dev](https://huggingface.co/black-forest-labs/FLUX.1-Krea-dev) - Hugging Face model card
72. [black-forest-labs/flux-3-action-base](https://huggingface.co/black-forest-labs/flux-3-action-base) - Hugging Face model card, September 2026