Black Forest Labs bets on open-source image models with sub-second FLUX.2 [klein]
On 15 January 2026, Black Forest Labs (BFL) — the Freiburg-based image-model company built around former Stable Diffusion lead authors Robin Rombach and Andreas Blattmann — released the FLUX.2 [klein] model family. Two variants (4B and 9B parameters), Apache 2.0 for the smaller, FLUX non-commercial license for the larger. The promise: sub-second inference on consumer GPUs starting at 13 GB VRAM. In the shadow of the release, a Series B closed in December 2025: $300M at a $3.25B valuation, co-led by Salesforce Ventures and Anjney Midha (a16z), with NVIDIA, General Catalyst and Temasek participating.
What changed in concrete terms
- Two models, two licenses: FLUX.2 [klein] 4B under Apache 2.0 (free for commercial use), 9B under the FLUX non-commercial license.
- Inference under 0.5 seconds for most prompts, BFL claims sub-second on capable hardware. Quantized FP8 and NVFP4 versions reduce VRAM usage by up to 55 percent.
- Hardware bar lowered: the 4B variant runs from 13 GB VRAM — workable on a single RTX 4060 Ti 16 GB or comparable. FLUX.1 [dev] previously needed closer to 24 GB+ for comfortable work.
- Series B in December 2025: $300M at $3.25B valuation. Salesforce Ventures and a16z co-lead, with NVIDIA, General Catalyst and Temasek.
- Distribution widens: FLUX.2 [dev] is now available on Cloudflare Workers AI — directly at the edge, no own GPU infrastructure needed.
What used to be the case
Since summer 2024, FLUX.1 [dev] and FLUX.1 [pro] had been the most credible open-weight argument against Midjourney and the proprietary image APIs from OpenAI and Google. The practical bottleneck, however, was hardware: anyone running FLUX.1 [dev] locally at acceptable speed needed at least a 24 GB card, with typical iteration in the range of several seconds per image. For agency workflows with “hundreds of variants in an hour”, that was the natural ceiling.
On the business side, BFL stood on a few large pillars: a $140M contract with Meta, partnerships with Adobe, Canva and Snap totalling around $300M in contract value, plus the direct API.
What’s the case now
1. Sub-second generation becomes the new default. With FLUX.2 [klein], the reference frame shifts: where you previously budgeted 4–6 seconds per image, the same workflow now runs at less than one second per image — on comparable hardware. That opens use cases that were not productive with FLUX.1: live iteration inside an editor, automated variant production, image generation as part of interactive sessions instead of batch jobs.
2. Apache 2.0 for the 4B variant lowers the entry bar significantly. Anyone who wants to build FLUX.2 [klein] 4B into their own products no longer hits license friction. The 9B variant remains under the FLUX non-commercial license — for commercial use of the 9B, the path is BFL’s paid API or an enterprise contract.
3. Consumer GPUs as target hardware. 13 GB VRAM for the 4B variant translates to: a 16 GB RTX 4060 Ti is enough. Quantized (FP8 / NVFP4) the memory footprint drops by up to 55 percent — bringing 8–12 GB cards into realistic territory. The move is strategic: BFL is not just addressing cloud customers, but studios and agencies running on-prem workflows.
Context
FLUX.2 [klein] is exactly what the open image-model scene needed — not yet another 70B parameter monster, but a noticeably smaller, faster model with an Apache 2.0 license for the commercial sub-variant. BFL’s strategy now reads clearly: open weights as a distribution channel, commercial contracts with Adobe, Canva, Meta and Cloudflare as the revenue base. The $300M Series B from December 2025 funds the gap in between — compute, data, talent.
Notable for the German AI narrative: BFL is the only German AI company that has stayed in the top tier of foundation-model providers in 2025/2026. Unlike Aleph Alpha, BFL did not chase ever-larger language models but picked a narrower, clearly bounded domain (images). And in that domain it has stayed financeable — a path other European specialists can learn from.
What you can do now
If you use image generation in agency workflows: test FLUX.2 [klein] 4B under Apache 2.0 locally. The sub-second latency unlocks a different iteration mode — no more “prompt and wait”, but interactive work on the image.
If you work cloud-first: FLUX.2 [dev] runs on Cloudflare Workers AI directly at the edge. That is the simplest path into productive pipelines without your own GPU infrastructure.
If you need to clear license questions: 4B = Apache 2.0 (commercial-free), 9B = FLUX non-commercial. For the 9B variant in commercial contexts, BFL’s enterprise API is the clean route — not the Hugging Face repo.
Entdecke mehr
Three Days with Claude Fable 5 — and Why Opus Stays My Daily Driver
Fable 5 fixed a bug in one shot that Opus failed at twice. I'm still not switching. My honest impressions from the trenches after three days.
GlossarMiniMax
MiniMax is an AI lab founded in Shanghai in 2021 whose open-weight M model family (LLMs) targets very long context windows and low running costs via sparse attention.
LexikonComparing AI models — who builds what and how to choose
The major model families in 2026 at a glance. Who builds Claude, GPT, Gemini, Llama, Mistral, DeepSeek, Qwen — and which model to pick when.