Black Forest Labs: FLUX.2 Klein 4B

What it is

FLUX.2 [klein] 4B is Black Forest Labs' Apache 2.0 licensed compact image model, released January 15, 2026 as the entry point of the Klein family, which BFL positions as its compact, minimal-latency image models. The name "klein" comes from the German word for "small," reflecting both the model's compact size and its minimal latency target — but BFL emphasizes it's "anything but limited," delivering performance in text-to-image, editing, and multi-reference generation typically reserved for much larger models.

Core specs

• It's a 4-billion-parameter rectified flow transformer capable of generating images from text and supporting multi-reference editing.
• It's distilled from Black Forest Labs' larger FLUX.2 base model, aiming to be more powerful than other models of similar size while retaining much of its larger "teacher" model's capability.
• The full inference pipeline targets under 500 milliseconds — sub-second generation.
• It fits in roughly 13GB VRAM and runs on consumer GPUs like the NVIDIA RTX 3090/4070 and above. BFL also released FP8 and NVFP4 versions built with NVIDIA for even smaller footprints on RTX hardware.

Capabilities

• Unified pipeline: It unifies text-to-image generation, image-to-image editing, and multi-reference generation in a single checkpoint.
• Image quality: It generates photorealistic images up to 4 megapixels, with accurate lighting, coherent spatial relationships, and readable text rendering.
• Text rendering: Unlike many small models that produce garbled text, Klein 4B generates clean, legible typography suitable for complex layouts, infographics, and UI mockups.
• Spatial/lighting realism: The model understands real-world lighting behavior — shadows, reflections, and perspective are handled correctly, reducing the typical "AI look."
• Editing: Supports high-resolution editing up to 4 megapixels while preserving detail and geometry rather than hallucinating new content during edits.

Licensing & positioning within the FLUX.2 family

• Klein 4B is Apache 2.0 licensed with no commercial-use restrictions — you can deploy it in commercial products, open-source projects, and redistribute derivatives freely. This distinguishes it from Klein 9B, which uses the FLUX Non-Commercial License, and FLUX.2 Pro, which is a proprietary API-only model.
• There's also a Klein 4B "Base" variant — an undistilled, full-capacity foundation model that preserves complete training signal for maximum flexibility, better suited to fine-tuning, LoRA training, and research, with higher output diversity than the distilled version but at some cost to raw speed.
• Benchmarks show Klein matches or exceeds Qwen's image quality at a fraction of the latency and VRAM, and outperforms Z-Image while still supporting both generation and multi-reference editing in one unified model.

Safety

BFL filtered pre-training data for NSFW content and known CSAM, and applied additional pre- and post-release mitigations aimed at reducing risks like CSAM and non-consensual intimate imagery generation before releasing the FLUX.2 family.

Best-fit use cases

Designed for real-time applications, creative iteration, and deployment on consumer hardware — think interactive editing tools, latency-critical production pipelines, local/edge deployment, and high-throughput services where cost-per-image matters more than squeezing out the absolute highest fidelity (that's what FLUX.2 Pro is for).

Want me to compare Klein 4B against Klein 9B or FLUX.2 Pro, or walk through how to run it locally/via API?