FLUX 3 Image Klein (5.6B open weights)
Confirmed
The takeaway
BFL’s own Elo tests put it first among open-weights image models and above closed Ideogram V4.5 and FLUX.2 [pro].
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 16 October 2026
Your AI and this story
- GPT-6 Astra168 days after its cutoff
- Claude Opus 5.5107 days after its cutoff
- Gemini 4 Argon198 days after its cutoff
- Grok 4.7137 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 107 days before it.
Key facts
- Size: 5.6B image transformer (text encoder Qwen3-VL-4B-Instruct, Apache-2.0, separate); base checkpoint (50 steps, guidance 3.5) for fine-tuning and a step-distilled checkpoint (16 steps, no guidance) for low latency; FP8, INT8 and NVFP4 builds
- Capabilities shared with FLUX 3 Image: native 4K generation, bounding boxes (0–1000 grid) to place, move, recolor, replace or remove elements; one model for text-to-image, single-reference editing and up to 10 reference images
- BFL’s internal Elo: first among open-weights models in text-to-image, multi-reference and editing; above closed Ideogram V4.5 and FLUX.2 [pro]; vs FLUX.2 [klein] 9B +91 Elo (text-to-image) and +70 (editing) with 38% fewer parameters; compared with Qwen-Image 2.1 (7B) and Krea 2 (12.9B)
- Speed (BFL): API median 4.54 s (text-to-image) and 7.18 s (editing) at 1 MP incl. queueing; locally ~4 s per 1 MP on an RTX 5090 and under 9 s on an RTX 4070 (distilled, INT8, ComfyUI); under 0.6 s on a GB300 with NVFP4; INT8 build runs on 8 GB VRAM
- Model card: ‘Do not use the model without a prompt upsampler’ (flux-3-image-klein-pup, a 9B model); peak memory on a B200 at 1024² without upsampler 31 GB (BF16) / 26 GB (FP8) / 25 GB (NVFP4)
- Licence: FLUX Kommunity License v1.0: commercial output generation for users under US$5M gross annualized revenue; otherwise a commercial licence from BFL; Hugging Face access gated (contact details)
- API: flux-3-image-klein on the BFL API; per-image price $0.008 (768²), $0.009 (1K), $0.011 (1.5K), $0.014 (2K), $0.036 (4K), same for generation and editing
What happened
Black Forest Labs (BFL) published FLUX 3 Image Klein on Oct 15, 2026 (the Hugging Face repository was created on Oct 13 and was trending on Oct 16). It is the small, open-weights member of the FLUX 3 family announced in July. The 5.6B model handles text-to-image, editing and multi-reference generation in one network and keeps two features of the larger FLUX 3 Image: native 4K output and bounding boxes that steer where elements go (they guide placement but are not hard masks).
BFL ships a base checkpoint for fine-tuning and LoRA training and a step-distilled one for fast generation, plus FP8, INT8 and NVFP4 builds. The model card insists on a separate 9B prompt upsampler, because short prompts give generic results. The blog says the INT8 build runs on GPUs with 8 GB of VRAM; the card’s B200 memory table (25–31 GB without the upsampler) measures a different setup, so local requirements depend on the build and the pipeline.
All quality and speed numbers are BFL’s own (internal Elo evaluations; API latency measured end to end, other models timed from the client, so BFL itself says the comparison is not like for like).
Why it matters
It brings FLUX 3-class editing and 4K output to consumer GPUs a day after Krea 3 and a month after Qwen-Image 2.1, and it uses the same “free under $5M revenue” licence pattern that open image models increasingly adopt instead of Apache-2.0 (FLUX.2 [klein] 4B was Apache-2.0).
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 4 primary
Primary
- Black Forest Labs: FLUX 3 Image Klein (blog)bfl.ai, official
- Hugging Face: black-forest-labs/flux-3-image-klein (model card)huggingface.co, code
- BFL API pricingdocs.bfl.ai, docs
- GitHub: reference codegithub.com, code
Changes
- Filed (found through Hugging Face trending)