FLUX.2
Black Forest Labs' frontier image model: 4MP output, up to 10 reference images, and a 32B open-weights release you can self-host.
Updated 2026-08-10
Yes, within limits. FLUX.2 runs a free tier with paid plans above it.
Listed pricing: FLUX.2 [pro] from ~$0.03/image via API; open weights for [dev] and [klein] — check site for current rates.
Overview
FLUX.2 is Black Forest Labs' flagship text-to-image and image-editing model, released in November 2025 and iterated since. BFL was founded by Robin Rombach and colleagues who built the original Stable Diffusion architecture, and FLUX.2 keeps their rectified-flow transformer core while swapping the text encoder for a Mistral-3 24B vision-language model instead of the usual CLIP/T5 stack. That change is the most consequential thing about the release: prompt following and legible in-image text — signage, packaging copy, UI labels — improved sharply over FLUX.1, which is where most diffusion models still fall over.
The lineup is four variants rather than the old pro/dev/schnell split. FLUX.2 [pro] is the hosted top tier. FLUX.2 [flex] is the same family with inference steps and guidance exposed as API parameters, so you can trade fidelity against cost per request instead of paying a flat rate. FLUX.2 [dev] is a 32B open-weights checkpoint on Hugging Face — genuinely downloadable, but released under a non-commercial license, so shipping commercial output from a self-hosted copy needs a paid BFL license. FLUX.2 [klein] is the size-distilled Apache-2.0 variant BFL announced alongside the launch for teams that need unrestricted terms. All of them handle up to 10 reference images in a single call and generate up to 4 megapixels natively, with depth, canny, inpainting and outpainting carried over as first-party controls.
The honest 2026 read: FLUX.2 is no longer the uncontested quality leader. Google's Nano Banana Pro generally wins on world-knowledge prompts and complex text layout, and Midjourney still owns aesthetics when you want a look rather than a likeness. What nobody else offers is weights. If your requirement is running a frontier-class image model on your own GPUs, fine-tuning it on proprietary product photography, or keeping inference inside your own network for compliance reasons, FLUX.2 is not the best option in its class — it is close to the only one.
Is FLUX.2 free?
Yes, within limits. FLUX.2 runs a free tier with paid plans above it.
What the free tier covers: FLUX.2 [klein] is Apache-2.0 and free to self-host commercially. FLUX.2 [dev] weights are free to download but the license is non-commercial — commercial output needs a paid license.
Listed pricing: FLUX.2 [pro] from ~$0.03/image via API; open weights for [dev] and [klein] — check site for current rates.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
What the free tier leaves out
Read straight off the plan list below. Vendors move features between tiers, so check the current split before you pay.
- FLUX.2 [pro] ~$0.03/image (verify current rate) Highest hosted quality, 4MP output, up to 10 reference images, commercial use included
- FLUX.2 [flex] Check website for current pricing Step count and guidance exposed as parameters — trade fidelity against cost per call
- FLUX.2 [dev] Check website for current pricing Hosted endpoint, or download the 32B weights free for non-commercial use; commercial self-hosting requires a paid BFL license
- Enterprise Custom Dedicated capacity, commercial self-hosting license, SLA and support
FLUX.2 pricing
| Plan | Price | What's included |
|---|---|---|
| FLUX.2 [pro] | ~$0.03/image (verify current rate) | Highest hosted quality, 4MP output, up to 10 reference images, commercial use included |
| FLUX.2 [flex] | Check website for current pricing | Step count and guidance exposed as parameters — trade fidelity against cost per call |
| FLUX.2 [dev] | Check website for current pricing | Hosted endpoint, or download the 32B weights free for non-commercial use; commercial self-hosting requires a paid BFL license |
| FLUX.2 [klein] | Free (Apache 2.0 weights) | Size-distilled open variant with unrestricted terms — the one to use if licensing matters more than peak quality |
| Enterprise | Custom | Dedicated capacity, commercial self-hosting license, SLA and support |
Highest hosted quality, 4MP output, up to 10 reference images, commercial use included
Step count and guidance exposed as parameters — trade fidelity against cost per call
Hosted endpoint, or download the 32B weights free for non-commercial use; commercial self-hosting requires a paid BFL license
Size-distilled open variant with unrestricted terms — the one to use if licensing matters more than peak quality
Dedicated capacity, commercial self-hosting license, SLA and support
3 plans carry no published number: FLUX.2 [flex], recorded as “Check website for current pricing”; FLUX.2 [dev], recorded as “Check website for current pricing”; Enterprise, recorded as “Custom”. Nothing is estimated in its place. The vendor's own pricing page is the only source for those figures.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
Is FLUX.2 worth it?
Worth it for Engineering and creative-ops teams building image generation into a product or pipeline — especially anyone who needs to self-host, fine-tune on proprietary assets, or hold a product or character consistent across thousands of renders.
You can test that on the free tier before paying anything. The recorded trade-offs are listed below, and any one of them can settle the question on its own.
The 8.8/10 AI Score is an editorial read of published capability, price and shipping pace. Nobody here has hands-on hours with FLUX.2. How we verify.
Worth it if
The strengths recorded against this entry.
- The only frontier-class image model shipping downloadable weights — you can quantize it, fine-tune it on proprietary assets, and keep inference inside your own network
- Up to 10 reference images per call, which is the difference between a consistent product catalogue and re-rolling until something matches
- In-image text rendering improved sharply over FLUX.1 — mockups, signage and packaging copy come out legible
- The [flex] tier makes cost per image a parameter you control rather than a fixed rate
- Depth, canny, inpainting and outpainting are first-party controls, not community extensions that break on each release
Not worth it if
Any one of these blocks your use case.
- 'Open weights' is doing heavy lifting — FLUX.2 [dev] is released under a non-commercial license, and teams routinely discover this after building on it
- 32B parameters is a serious deployment: full-precision inference wants far more VRAM than a single consumer card, and 24GB setups need FP8/GGUF quantization plus CPU offloading
- Google's Nano Banana Pro generally beats it on world-knowledge prompts and complex text layout, so FLUX.2 is no longer the default answer for one-shot quality
- Still tuned for photorealism and control over style — Midjourney remains the better pick when you want an aesthetic rather than a likeness
What sets FLUX.2 apart
- Ships downloadable weights at frontier quality — Nano Banana Pro, GPT Image and Midjourney are all API or app only
- Accepts up to 10 reference images in a single generation for character and product consistency
- The [flex] tier exposes inference steps and guidance as API parameters, so per-image cost is tunable rather than fixed
- Uses a Mistral-3 24B vision-language model as its text encoder rather than the conventional CLIP/T5 stack
Key features
Multi-Reference Editing
Condition a single generation on up to 10 reference images at once. This is the practical mechanism for holding a character, product, or brand style consistent across a large batch rather than re-rolling until something matches.
4MP Native Output
Generates up to 4 megapixels straight from the model, no separate upscaling pass. In-image text rendering is substantially better than FLUX.1, which makes mockups, packaging, and signage usable rather than decorative.
Open Weights, Two Licenses
FLUX.2 [dev] is a 32B checkpoint on Hugging Face under a non-commercial license; FLUX.2 [klein] is the size-distilled Apache-2.0 variant. Read the license before you build a business on either — they are not interchangeable.
Tunable Inference via [flex]
The [flex] tier exposes step count and guidance as request parameters, so cost per image is something you dial rather than accept. Depth, canny, inpainting and outpainting ship as first-party controls across the family.
How it compares
| Tool | Best for | Pricing | Score |
|---|---|---|---|
| FLUX.2 | Engineering and creative-ops teams building image generation into a product or pipeline — especially anyone who needs to self-host, fine-tune on proprietary assets, or hold a product or character consistent across thousands of renders. | FLUX.2 [pro] from ~$0.03/image via API; open weights for [dev] and [klein] — check site for current rates | 8.8/10 |
| Imagine Image 2.0 vs Imagine Image 2.0 → | Social and marketing creatives who need one visual in three aspect ratios with the background cut and one object fixed, without paying for a generator plus a background remover plus a reframing tool. | Free with a Grok account; SuperGrok $30/mo, Heavy $300/mo; API ~$0.06/image | 9/10 |
| Midjourney vs Midjourney → | Concept artists, art directors and illustrators who want a striking, coherent house style out of the box and don't need to call the model from code. | From $10/mo (no free tier) | 8.8/10 |
| Reve Image vs Reve Image → | Designers and marketing teams producing posters, ad creative and UI mockups where a misspelled word inside the image or a misplaced element makes the asset unusable. | Free tier + Lite $7.99/mo, Pro $19.99/mo (June 2026 tiers — check site) | 8.5/10 |
Compare head-to-head
Related reading
Muse Glimmer and the 30B Open-Weights Ceiling
Meta released Muse Glimmer's 30B weights under Apache 2.0, landing in the same consumer-GPU size class as every other recent open model release.
OpenAI Astra's Cyber Pause and the Release Rules
OpenAI flagged Astra as critical on cybersecurity and paused non-essential work on it. What the hold covers and what nobody outside OpenAI can check.
ChatGPT Slider and Think Button: How to Use Both
OpenAI put reasoning depth under your control in ChatGPT. Which tier gets which control, which setting fits which task, and when to leave both alone.
Ready to try FLUX.2?
Head to the official site to start with FLUX.2 — pricing and plans are listed above.
Visit FLUX.2