FLUX 3 Image Klein 已发布：56 亿参数、8 GB 显存，以及许可证中的 500 万美元条款

Black Forest Labs 于 2026 年 10 月 15 日发布了FLUX 3 Image Klein：这是一个拥有 5.6B 参数的开源权重图像模型，同时也在该公司自己的 API 上以按量计费的端点形式运行，端点名为 flux-3-image-klein/，起价为每 1K 张图像 0.009 美元。它隶属于规模更大的 FLUX 3 Image，并继承了该模型的主要卖点——原生 4K 输出、最多十张参考图像，以及用于放置或编辑元素的边界框——而每 1K 张图像的价格大约只有其五分之一，体量也仅为其一小部分。这些都是发布时的宣称，可以对照权重、许可证和价格表来核实。尚无法核实的是那个决定大多数购买决策的宣称：一个 5.6B 模型在质量上胜过规模更大的开源权重竞争对手。

发布版本里的每一个质量数字都来自Black Forest Labs内部。公司之外没有任何人发布过基准测试，公开排行榜上也还没有Klein的条目。对于一款问世仅一天的模型来说，这很正常，而且值得一开始就说明，因为本文其余部分异常具体——文件大小、每张图像的费率、许可证阈值——而质量的问题与这一切是分开的。

10月15日实际发布的是什么？

公告标题为FLUX 3 Image Klein：开放权重，5.6B 参数，其中列举的是一次完整发布，而非仅仅投放单个检查点：

• Two checkpoints. Klein Base is undistilled bfloat16, runs 50 steps at guidance 3.5, and is the one meant for fine-tuning, LoRA training and research. Klein distilled is bfloat16 at 16 steps with no guidance, and is the checkpoint behind the B​FL API.

• Three quantisation formats. FP8 and NVFP4 builds ship alongside the BF16 transformer in the main repository, with an INT8 build developed with N​VIDIA support. The ComfyUI repackaging carries INT8 versions of both the transformer and the text encoder.

• 一个模型，胜任三项任务。文生图、单参考图编辑和多参考图合成均通过同一套权重运行。参考图用于锚定内容，而非锁定结构，且模型最多可接受十张参考图。

• Bounding boxes on a 0–1000 grid. Each element gets a box and a description, written as [top, left, bottom, right]/, and the box can either place a new element or confine an edit to a region. B​FL is explicit in the model card that boxes guide placement and scale rather than constraining pixels to a hard mask — small requested elements may not appear at all, and an element can extend past its box.

• Day-one runtime support in ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, plus LoRA training through Ostris's AI Toolkit and recipes in N​VIDIA NeMo AutoModel. The model was trained on N​VIDIA GB200 NVL72.

这些仓库在公告发布之前就已上线——基础版和蒸馏版仓库创建于10月13日，ComfyUI 仓库和提示词上采样器仓库创建于10月14日——而发布博文、API 端点和定价页面则都在15日才跟进。img src="2.png" alt="无头浏览器截取的 Hugging Face 上 black-forest-labs/flux-3-image-klein 仓库页面，展示了许可证接受门槛、flux-kommunity-license 徽章、标注为 6B 参数且张量类型为 BF16 的模型规模、上月 3,241 次下载、219 个赞，以及列出适配器和微调模型的模型树，界面为英文，截取于 2026 年 10 月 16 日。" />

8 GB 的说法，以及是谁测量的

"Runs on a single consumer GPU" is the sentence that will get quoted most, and B​FL does give it a number: with the INT8 build, Klein runs on GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers. The generator weights themselves are 11.2 GB in BF16 and 5.6 GB in FP8, per the blog post.

The speed figures come from N​VIDIA, not from Black Forest Labs, and they are specific about the setup: the step-distilled model with the INT8 build produces a 1 MP image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070, measured in ComfyUI. On a GB300 running NVFP4, the same distilled model lands under 0.6 seconds. Serving a LoRA adds roughly 5% to generation time.

One structural caveat sits in the latency chart B​FL published alongside the numbers, and the post discloses it itself: Klein was timed on B​FL's own servers, while the compared models — Q​wen-Image 2.1, FLUX.2 [pro], I​deogram V4.5 and K​rea 2 — were timed from the client, which includes polling and network time. The RTX figures are N​VIDIA's ComfyUI measurements. Neither set is an independent benchmark of quality.

质量声明的真实立场，实话实说

B​FL's position is that Klein "ranks first in text-to-image, multi-reference generation and image editing, ahead of larger open weights models" in its internal evaluations, and that it rates above the closed models in its comparison. Against its own previous small model, FLUX.2 [klein] 9B, B​FL reports a gain of 91 Elo in text-to-image and 70 in image editing with 38% fewer parameters.

Those are vendor-reported and unreproduced. Two details make them easier to weigh. First, the comparison set: 7B for Q​wen-Image 2.1 and 12.9B for K​rea 2, with parameter counts taken as each publisher reports them and text encoders excluded from every count — the same accounting that makes Klein "5.6B" while shipping an 8.9 GB Qwen3-VL-4B-Instruct text encoder next to it. Second, the model B​FL did not put in the chart: FLUX 3 Image, the larger sibling, is absent from its own smaller model's comparison. There is no published number anywhere saying how far Klein sits below it.

Independent boards have not caught up. As of 16 October, Artificial Analysis carries an entry for FLUX 3 Image on its text-to-image board and none for Klein, so there is no third-party Elo for the model at all. The first number that does not come from B​FL will be the interesting one.

成本是多少，以及曲线如何弯曲

The API rate is the sharpest thing in the release. On B​FL's published image pricing, the distilled Klein checkpoint is:

• 768 × 768 — 每张图片 $0.008

• 1K — 每张图片 $0.009

• 1.5K — 每张图片 $0.011

• 2K — 每张图片 $0.014

• 4K — 每张图片 $0.036

同一张表将 FLUX 3 Image 在五个档位的价格定为 $0.041、$0.048、$0.07、$0.10 和 $0.607。这意味着在 1K 下存在 5.3× 的差距，在 4K 下大约有 17× 的差距，而这正是 Klein 的全部商业论据：如果你的工作负载是批量进行的 4K 文本到图像，那么更小的模型并不是你勉强接受的降级，而是一个会按数量级变化的成本项。

注意曲线的形状，而不是入门价格。Klein 最便宜的档位是 $0.008，但费率一直到 2K 都基本持平，到了 4K 则跳涨 2.6 倍；而 FLUX 3 Image 的涨幅超过 6 倍。如果你的输出是 1K，两个模型的价格足够接近，该由质量来决定取舍。如果是 4K，它们就完全不在同一个讨论范围内了。

许可证决定了你是否真的能发布。

Weights under the FLUX Kommunity License v1.0 are free for non-commercial use, and the licence grants a commercial-output exception to "Qualifying Users" — individuals and companies with under US$5M in gross annualized revenue. Above that line, deploying Klein commercially means a self-hosted licence from B​FL or going through the API.

仔细阅读措辞，因为对它的两种概述在范围上并不一致。发布帖将其描述为“面向个人以及年收入低于 500 万美元的公司免费商用”。模型卡则把同一例外表述为涵盖合格用户（Qualifying Users）的商业输出生成，并补充说它“并非对商业微调或分发的一揽子许可”。对于一个计划在专有数据上微调 Klein 并将所得模型作为产品的一部分发布的团队来说，发布帖并不是回答该问题的文件。完整许可证文本才是。

在围绕 Klein 设计数据管道之前，还有一条值得了解的条款：该许可证限制将模型、其衍生品或其输出用于改进另一个执行与 FLUX 模型类似功能的 AI 模型。为训练竞争性图像模型而生成合成数据，正是该条款所针对的情形。

你绝不能跳过的一件事

Klein does not take short prompts. Both the model card and B​FL's page say the same thing in different words: it expects full scene captions covering framing, subject, background and lighting, and a one-line request sent as typed returns generic output. The reference implementation therefore runs a prompt upsampler before every generation — text-to-image and editing alike — and the model card is blunt about it: Do not use the model without a prompt upsampler.

The cost of that is real. B​FL's upsampler, flux-3-image-klein-pup/, is a separate 18.8 GB download in BF16 and adds about 28 GB of GPU memory on an 80 GB card when it runs alongside the generator. The documented opt-out is to supply your own: any LLM can take its place if you feed it B​FL's system prompt and pass the caption it returns. For anyone planning an on-device or low-VRAM deployment, the upsampler is the piece that quietly breaks the 8 GB budget, and the fix is to run captioning somewhere else and send Klein the finished caption with the prompt-as-typed flag.

开始处理，以及接下来会发生什么

权重托管在 Hugging Face 上，需先接受许可协议才能访问，同一组织下还提供参考推理代码和一个用于边界框控制的 Gradio 演示。托管端点就是厂商自己的 API。Klein 目前不在 OrcaRouter 的目录中，所以我们这里没有什么可供路由的，与其暗示相反，我们更愿意直说。OrcaRouter 确实能为你提供的，是真实图像流水线的另一半：被收录在目录中的托管图像模型，可与其他 200 多个模型共用一个密钥，供应商的标价按 0% 加价原样透传，并在某个供应商出现波动时自动故障转移。最终用 Klein 跑批量 4K 工作、同时调用托管模型处理交互路径的团队，正是在做单一端点本就为此设计的那种拆分。

What to watch over the next fortnight is narrower than usual. Three things would settle the open questions: an independent Elo for Klein, ideally on the same board as the Q​wen and K​rea models B​FL benchmarked itself against; a reported memory figure for the INT8 build with the upsampler actually in the loop, rather than the generator alone; and a first look at what fine-tunes do to the model's behaviour, given that LoRA training was supported on day one and the licence has opinions about what those fine-tunes may be used for. Until the first of those lands, treat every quality comparison — including the one in this article — as a vendor's own chart.

基准测试会过时，价格会变动，但失败记录不会：这一类模型背后有据可查的 AI 事故与智能体故障，均收录于OrcaRouter AI 事故档案库，其中每条记录都带有各自的来源、严重程度、置信度评级和争议标记。