Open Source
Open Source

5.6B Beats 9B: FLUX 3 Klein Opens Its Weights

Black Forest Labs put FLUX 3 Image Klein's weights on Hugging Face on 2026-10-15, at huggingface.co/black-forest-labs/flux-3-image-klein, alongside a fine-tunable base repository and a gated ComfyUI repository. The image generator is a 5.6B flow-matching transformer built, per the official description, from per-mode early blocks plus 28 joint blocks, with a Qwen3-VL-4B text encoder and the existing FLUX.2 VAE; parameter counts are each publisher's own and exclude text encoders. The headline is that the smaller generation wins: against its own predecessor FLUX.2 [klein] 9B it gains 91 Elo in text-to-image and 70 in image editing with 38 percent fewer parameters, and in BFL's internal evaluations it ranks first in text-to-image, multi-reference generation and image editing, ahead of the closed Ideogram V4.5 and its own FLUX.2 [pro], with Qwen-Image 2.1 at 7B and Krea 2 at 12.9B as the open comparisons. Krea 2 does not support image editing and so is absent from that comparison - a detail worth keeping, or three firsts get misread as sweeping every rival in all three. Capabilities: native 4K with dimensions in multiples of 16, pixel-perfect editing where untouched regions are preserved exactly rather than approximately repainted, bounding boxes that place elements or confine edits to a region, up to 10 reference images for character consistency and style transfer without retraining, and an optional Qwen3.5-9B prompt upsampler. Memory comes in three tiers: 11.2GB in BF16, 5.6GB in FP8, and an INT8 build that runs on NVIDIA GeForce RTX cards with as little as 8GB of VRAM through ComfyUI or Diffusers. Vendor-reported speed puts a 1MP image at about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 with the distilled INT8 build, under 0.6 seconds in NVFP4 on an NVIDIA GB300, with a served LoRA adding roughly 5 percent. The licence is the FLUX Kommunity License: free commercial use for individuals and companies under 5 million dollars in annualized revenue, free research and non-commercial use for everyone, and self-hosted licensing or the API beyond that threshold, with API pricing from 0.009 dollars per 1K image, the LICENSE text being the final authority. Three caveats belong side by side: every quality figure is BFL's internal evaluation with no independent reproduction yet; the API latency comparison is not like-for-like, since BFL timed its own servers while rivals were timed from the client including polling and network; and the 8GB claim is not tied to a specific variant, output size or generation time, with the published speed charts not testing an 8GB configuration or 4K output at all. Day-one support covers ComfyUI via PR 17064 plus four official templates, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, with LoRA training through Ostris AI Toolkit and NVIDIA NeMo AutoModel.

Published October 17, 202610 min read
<!-- flux-3-image-klein-opensource-resource | open-source | 5.6B Beats 9B: FLUX 3 Klein Opens Its Weights -->

On October 15, 2026, Black Forest Labs (BFL) published the weights for FLUX 3 Image Klein at black-forest-labs/flux-3-image-klein. It is the second open-weight release in the FLUX 3 family, after FLUX 3 Action (7B, robot control). Three repositories: the distilled main repo, a base repo for fine-tuning, and a ComfyUI repo that is gated, so you must be logged in and have accepted the license before a download succeeds.

Repository: huggingface.co/black-forest-labs/flux-3-image-klein

This piece answers three questions: how a 5.6B model beats a 9B one, how much substance sits behind the 8GB VRAM claim, and what the FLUX Kommunity License permits. Every figure below is BFL's own or a third-party relay; we have not run independent benchmarks.

1. How a 5.6B model beats a 9B model by 91 Elo

Klein's image generator is a 5.6B flow matching transformer: per-mode early blocks plus 28 joint blocks, a Qwen3-VL-4B text encoder and the FLUX.2 VAE. Parameter counts are vendor-reported and exclude the text encoder.

The comparison that matters is against BFL's own predecessor, FLUX.2 [klein] 9B: plus 91 Elo on text-to-image, plus 70 on image editing, with 38% fewer parameters. Those three numbers are the story, because "the smaller generation is stronger" is not the inevitable version of "the new generation is stronger." In BFL's internal evaluation Klein ranks first across text-to-image, multi-reference generation and image editing, ahead of closed models including Ideogram V4.5 and BFL's own FLUX.2 [pro]. The open-weight comparisons are Qwen-Image 2.1 at 7B and Krea 2 at 12.9B. Krea 2 does not support editing, so it was absent from that comparison, which is why the result is not a clean sweep of every opponent in all three categories.

Why the smaller model can overtake the larger one, BFL does not itemize. Joint blocks dominate the transformer, the encoder moved to Qwen3-VL-4B, the VAE carries over from FLUX.2, and how much Elo each change contributes is not something an outside observer can attribute.

Elo needs reading carefully too: pairwise human votes, and a larger gap means stronger preference agreement, but conversion to a win rate depends on rating pool, voter count and prompt difficulty, none published. Read plus 91 as a clear preference gap on BFL's prompt set, not 91 points stronger everywhere.

Family bookkeeping, briefly: FLUX 3 has two open-weight members, Action for robot control and Image Klein for generation and editing. Klein ships as base for fine-tuning and step-distilled for few-step output.

2. One set of weights, three jobs: 4K, pixel-perfect edits, box-guided placement

Native 4K. Output tops out at 4K. "Native" means generating at high resolution directly rather than upscaling afterward. Print work and posters need it; social posts do not, and pay in time and VRAM.

Pixel-perfect editing. Regions you do not edit are preserved exactly, the official word being "exactly," not "approximately." Conventional generative editing regenerates the whole frame, so untouched areas shift in pixel values, noise and text edges, and after a few rounds the image goes soft. This is the hard requirement for e-commerce and product shots.

Bounding box control. Specify where an element lands, or confine an edit to a boxed region. Composition control and scope control respectively, and the box is more reliable than prose when one frame holds several people or products.

Both output dimensions must be multiples of 16, so existing assets cannot be fed in as-is: crop or pad at the pipeline entrance.

Multi-reference generation. Up to 10 reference images for character consistency and style transfer, with no retraining. Ten is generous for this tier. In practice that means one character sheet: generate a few angles once, then reuse them as references instead of retraining a LoRA for every new pose.

Optional prompt upsampler. A Qwen3.5-9B-based enhancer that expands a short instruction into a full scene description. It is optional, so the model runs without it, but short prompts land less reliably.

One model covering all three jobs means one set of weights and three entry points instead of three pipelines. The cost is more entry points, and first-time users will pick the wrong template at least once.

3. The 8GB line: what the four precision builds actually mean

PrecisionWeight footprintOfficial positioning
BF1611.2 GBHigh VRAM, first choice for fine-tuning
FP85.6 GBMid-range VRAM
INT8Runs on RTX cards with as little as 8GBConsumer laptops, mid-tier cards
NVFP4No consumer guidance givenData-center throughput

The line that travels furthest: the INT8 build runs on NVIDIA GeForce RTX cards with as little as 8GB of VRAM, through ComfyUI or Diffusers. That moves local 4K generation and pixel-level editing out of high-VRAM territory and onto a laptop dGPU.

On speed (vendor figures, not independently reproduced): distilled plus INT8 at 1MP takes about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070; NVFP4 on an NVIDIA GB300 comes in under 0.6 seconds; a LoRA adds roughly 5% to generation time.

The NVFP4 row is the one to ignore on a desktop card. It exists for data-center throughput, where a GB300 serves many requests rather than one image at a time.

BFL publishes no quality delta between the precision builds. Lower precision saves memory, and what you save is usually paid back in image quality or stability; only your own A/B run on one prompt set will tell you how much. For production, test the exact build you plan to ship rather than the BF16 reference.

Now the brake. The 8GB claim is tied to no variant, no output size and no generation time. Distilled or base, 1MP or 4K, 30 seconds or three minutes, the announcement does not say. The official speed chart tests neither an 8GB configuration nor 4K output. So "runs on 8GB" means "loads and completes a run," not "comfortably produces 4K." Budget for 1MP: the 8GB figure is a floor for loading, not a target for throughput.

4. Three caveats worth remembering

First, all of it is BFL's internal evaluation. The plus 91 Elo, the three first-place finishes, the win over Krea 2: all vendor-run, with no independent replication yet. Third-party outlet ClawdBytes says plainly that these figures are not independent verdicts. Prompt set, rater pool and baselines were all the vendor's choice, so the direction is useful and the magnitude is not.

Second, the API latency comparison is not apples to apples. BFL's medians of 4.54 seconds for text-to-image and 7.18 seconds for editing are server-side timings on BFL's own infrastructure. Competitor figures are client-side, which includes polling and network overhead.

Third, the 8GB claim and the speed chart leave the same gap. No variant, no output size, no generation time, and no 8GB or 4K run in the chart.

All three point one way: the direction of this release, smaller models overtaking larger ones with a lower VRAM floor, is credible, and the numbers should wait for replication. Before you cite a vendor table in a decision, ask which prompt set, which sampler and which precision it used; the published chart rarely answers all three. A batch of your own images on your own card beats an Elo table.

5. Licensing: the $5 million line

FLUX 3 Image Klein ships under the FLUX Kommunity License (v1.0), neither Apache nor MIT. Three points are confirmed:

  1. Individuals and companies with annualized total revenue below $5 million may use it commercially at no charge.
  2. Research and non-commercial use is free for everyone.
  3. Organizations above that line need a self-hosted commercial license or must go through BFL's API.

Per the same license as published for FLUX 3 Action, the threshold is computed on annual total revenue plus affiliates on a consolidated basis, not per entity or per product line. If the parent is over the line, a small subsidiary does not get a pass. The license also attaches conditions to output moderation and filtering and to redistribution, so this is not "download it and ship it anywhere."

To be explicit: we did not capture the full text of Klein's own LICENSE, because the HF repo is gated and reading it requires accepting the terms. Clause detail is therefore deferred to the LICENSE file in the repository, and this article will not extrapolate. Treat these three points as a screen, not legal advice. What a self-hosted commercial license costs is also unpublished.

Qwen-Image 2.1 has the opposite problem: an incomplete Apache license, as our earlier piece covered. Klein is clear but gated by revenue. Neither is "use it however you like."

6. Ecosystem: what the day-one list tells you

Launch-day support covers ComfyUI (PR #17064, plus four official templates), Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni. LoRA training runs through Ostris AI Toolkit and NVIDIA NeMo AutoModel, both with ready-made recipes; training hardware is listed as NVIDIA GB200 NVL72. The two LoRA routes matter because fine-tuning is where open weights usually earn their keep.

TensorRT-LLM, SGLang Diffusion and vLLM-Omni are inference-serving stacks rather than hobbyist tools, which suggests BFL was building for batch generation and per-request billing from day one. Day-one support also shortens the path from downloaded weights to a working service: no waiting on a community port. The ComfyUI repo shipping INT8 weights directly says quantization and consumer hardware were first-class targets, not an afterthought patch.

The Diffusers path is where programmability lives: reference inference code and a Gradio demo with bounding box control, for embedding generation into your own service. Class names and parameters belong in the repository README, not in this article.

On safety, BFL cites pre-training data filtering plus targeted post-training, an internal red team before release, and an external evaluation with partner Cinder. Commercial API pricing starts at $0.009 per 1K-resolution image (vendor figure; full tiers per official announcement).

7. Who should pull it now, and who should wait

Pull it now if:

  • You have an RTX card with 8GB or more and want local image editing. Start at 1MP.
  • You are a small team needing character consistency or style transfer. Ten reference images and one set of weights beats three pipelines.
  • You want to fine-tune. The base repo plus the Ostris and NeMo recipes is a working path.
  • Your company is under $5 million in annual revenue and needs commercial output.

Wait if:

  • You need reliable 4K in production. BFL has not tested 8GB with 4K.
  • You are over the revenue line and do not want the API. A self-hosted license is a negotiation; model the volume first.
  • You only trust independent benchmarks. Every ranking available today is vendor-sourced.
  • Your budget only covers a cloud API and your volume is low: $0.009 per image usually beats electricity plus hardware at small scale.

Deployment specifics, including ComfyUI templates, the file list and two documented pitfalls, live in a separate local deployment SOP.

The value here is structural. One 5.6B model covers generation, editing and multi-reference work, drops the floor to 8GB, and comes with a license that has a clear revenue line. Whether it truly beats a 12.9B competitor needs independent testing. Confirm three things first: whether your card runs it, whether your revenue sits under the line, and whether the license covers how you intend to ship.

FAQ

Q1: Is FLUX 3 Image Klein open source or open weight?

A1: The weights are publicly downloadable, but the FLUX Kommunity License carries a revenue threshold and redistribution conditions and is not a general-purpose open-source license like Apache or MIT. "Open weight" is the more accurate label; for a strict determination, read the LICENSE file.

Q2: Is 5.6B really stronger than 9B?

A2: By BFL's internal evaluation: plus 91 Elo on text-to-image, plus 70 on editing, with 38% fewer parameters. Those are vendor figures with no independent replication. The direction is credible; the deltas need third-party confirmation.

Q3: Can 8GB of VRAM handle 4K?

A3: BFL states only that the INT8 build runs on RTX cards with as little as 8GB, tying it to no variant, output size or generation time, and the speed chart tests neither 8GB nor 4K. Plan for 1MP at 8GB and add VRAM for 4K.

Q4: Does crossing $5 million in revenue mean I cannot use it?

A4: It means you cannot use the free commercial terms. Above the threshold you need a self-hosted commercial license or BFL's API, computed on annual total revenue plus affiliates on a consolidated basis; confirm against the LICENSE text.

Q5: Where do I download it, and why is the ComfyUI repo blocked?

A5: The main repo is huggingface.co/black-forest-labs/flux-3-image-klein, and the fine-tuning version adds a -base suffix. The ComfyUI repo is gated: log into Hugging Face and accept the license on the page, or requests return 401.

This article is AI-assisted and human-edited. Last updated: 2026-10-17

FAQ

Is FLUX 3 Image Klein open source or open weight?
The weights are publicly downloadable, but the FLUX Kommunity License carries a revenue threshold and redistribution conditions and is not a general-purpose open-source license like Apache or MIT. "Open weight" is the more accurate label; for a strict determination, read the LICENSE file.
Is 5.6B really stronger than 9B?
By BFL's internal evaluation: plus 91 Elo on text-to-image, plus 70 on editing, with 38% fewer parameters. Those are vendor figures with no independent replication. The direction is credible; the deltas need third-party confirmation.
Can 8GB of VRAM handle 4K?
BFL states only that the INT8 build runs on RTX cards with as little as 8GB, tying it to no variant, output size or generation time, and the speed chart tests neither 8GB nor 4K. Plan for 1MP at 8GB and add VRAM for 4K.
Does crossing $5 million in revenue mean I cannot use it?
It means you cannot use the free commercial terms. Above the threshold you need a self-hosted commercial license or BFL's API, computed on annual total revenue plus affiliates on a consolidated basis; confirm against the LICENSE text.
Where do I download it, and why is the ComfyUI repo blocked?
The main repo is huggingface.co/black-forest-labs/flux-3-image-klein, and the fine-tuning version adds a -base suffix. The ComfyUI repo is gated: log into Hugging Face and accept the license on the page, or requests return 401.

Related

Open Source

Univer Office SDK: Agents Edit Cells Without an AI SDK

Only what the dream-num/univer README, LICENSE and GitHub API fields support. Apache-2.0 covers the Web SDK core architecture and engine implementations; commercial plugins, the Server SDK and commercial AI SDK extensions are distributed under separate licences. Facade APIs such as FWorkbook, FWorksheet, FRange, getRange and setValues let an application or Agents Edit Cells directly with formula recalculation, while workbook.save() persists a JSON snapshot rather than XLSX import or export. No AI SDK is required anywhere in that path. Stars are cumulative GitHub API totals, not single-day growth.

Nov 16, 20269 min read
Open Source

llama.cpp and ggml Ship v1.0.0: Deprecated APIs Removed

Reads only what the GitHub API fields, the two v1.0.0 release notes, Discussion 1579, LICENSE and README of the ggml-org llama.cpp and ggml repositories can verify. Both v1.0.0 tags landed on November 13, 2026. It separates semantic versioning, the rule that a release tag requires the internal ggml copy to match the released one, the split between b-numbered nightlies and vX.Y.Z releases, and the 35 removed functions, RPC v9 and MIT notice-retention and AS IS terms. The August 2026 versioning discussion is not presented as a November announcement, and no figure from the unretrieved HF blog is quoted.

Nov 15, 20269 min read