FLUX.2 [klein]: Towards Interactive Visual Intelligence

TL;DR

Black Forest Labs’ fastest image-generation family: a 9B and a 4B rectified-flow transformer, step-distilled to 4 inference steps, that do text-to-image, single-reference editing, and multi-reference editing in one unified compact architecture. The 9B distilled checkpoint claims sub-0.5s end-to-end inference on modern hardware while matching or exceeding models 5× its size; the 4B runs in ~13GB VRAM on an RTX 3090/4070. Shipped as a six-checkpoint release (9B / 4B × distilled / base, plus FP8 and NVFP4 quantized variants developed with NVIDIA), with 4B variants under Apache 2.0 and 9B variants under the FLUX Non-Commercial License. Positioned explicitly as a step toward “interactive visual intelligence” — real-time design tools, agentic visual reasoning, interactive content creation — and is the model powering the realtime demos on fal that prompted this filing.

Key claims

The 9B distilled model is step-distilled to 4 inference steps and pairs a 9B flow transformer with an 8B Qwen3 text embedder [§The FLUX.2 [klein] Model Family].
Sub-second end-to-end inference: generate or edit images in under 0.5s on modern hardware [§What’s New].
The 4B variant fits in ~13GB VRAM and runs on consumer GPUs (RTX 3090/4070 and above) [§What’s New].
Generation and editing — text-to-image, single-reference editing, multi-reference editing — are unified in a single architecture rather than separate checkpoints [§What’s New].
The Base 9B / Base 4B checkpoints are undistilled and retain higher output diversity, intended for fine-tuning, LoRA training, and research [§The FLUX.2 [klein] Model Family].
FP8 quantization gives up to 1.6× speedup and up to 40% VRAM reduction; NVFP4 gives up to 2.7× speedup and up to 55% VRAM reduction (measured on RTX 5080/5090, T2I at 1024×1024) [§Quantized versions].
On the reported Elo-vs-latency and Elo-vs-VRAM frontier, FLUX.2 [klein] matches or exceeds Qwen’s quality at a fraction of latency and VRAM, and outperforms Z-Image while supporting both T2I and multi-reference editing in one model [§Performance Analysis].
Licensing: 4B and 4B Base variants are Apache 2.0; 9B and 9B Base variants are under the FLUX Non-Commercial License (the renamed FLUX [dev] NCL) [§The FLUX.2 [klein] Model Family].

Method

FLUX.2 [klein] is a rectified-flow transformer family in two sizes (4B and 9B), conditioned on text via a frozen 8B Qwen3 text embedder (continuing the trend of frozen-LLM-as-text-encoder seen in LTX-2 and other recent T2I/T2V models). The 9B and 4B distilled checkpoints are step-distilled from the corresponding Base checkpoints to 4 inference steps; the Base variants preserve the full training signal and trade speed for output diversity and fine-tunability. The architecture unifies text-to-image, single-reference image editing, and multi-reference editing in one model — the same checkpoint handles all three modes depending on what conditioning is supplied at inference.

Six total checkpoints in the release: 9B distilled, 4B distilled, 9B Base, 4B Base, plus FP8 and NVFP4 quantized versions of all variants (developed in collaboration with NVIDIA, targeting RTX inference). Available via the BFL API, via fal.ai (which exposes a websocket-based realtime endpoint — the basis for the pencil-autocomplete demo that prompted this filing), and as open weights on Hugging Face / GitHub.

Results

The blog post reports Elo-vs-latency and Elo-vs-VRAM Pareto frontiers across T2I, I2I single-reference, and multi-reference tasks, with speed measured on a GB200 in bf16. Headline comparisons:

Matches or exceeds Qwen’s image-generation quality at a fraction of the latency and VRAM [§Performance Analysis].
Outperforms Z-Image while additionally supporting multi-reference editing [§Performance Analysis].
9B distilled: sub-0.5s end-to-end inference [§What’s New, §The FLUX.2 [klein] Model Family].
4B distilled: fits in ~13GB VRAM, runs on RTX 3090/4070-class hardware [§What’s New].
FP8: up to 1.6× faster, up to 40% less VRAM vs bf16 [§Quantized versions].
NVFP4: up to 2.7× faster, up to 55% less VRAM vs bf16 [§Quantized versions].

Exact Elo deltas, FID/CLIP numbers, and per-task latencies are reported on the linked plots but not enumerated in the blog body.

Why it’s interesting

Two reasons for Luma. First, FLUX.2 [klein] is now the open reference point for the sub-second / consumer-GPU image-generation envelope — the latency-quality frontier any Luma image-generation work will be compared against and the obvious initial backbone for any realtime-creative-tooling prototype the team wants to ship quickly. Second, the model is what powers the fal realtime endpoints driving demos like @tomasproc’s “pencil autocomplete” — the specific UX gray flagged in the originating Slack post. The combination of unified generation+editing in one checkpoint, sub-second latency on commodity hardware, and an open Apache-2.0 4B variant makes this the most direct off-the-shelf substrate for that workflow. It also fits the Open foundation-model releases pattern the wiki has been tracking — six checkpoints, multiple quantization formats, joint hardware-partner collaboration on quantization — and is the first image-generation entry to that concept page explicitly positioned around interactive real-time inference rather than maximum quality at any latency.