On October 15, 2026, Black Forest Labs released FLUX 3 Image Klein as open weights, and BFL’s release documents an 8GB VRAM path for the INT8 build while publishing no VRAM guidance for the rivals it names, the 7B Qwen-Image 2.1 and the 12.9B Krea 2. Per BFL’s release post, the 5.6B-parameter model’s generator weights need 11.2 GB in BF16 and 5.6 GB in FP8, and “With the INT8 build, FLUX 3 Image Klein runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM, through ComfyUI or Diffusers.”
That answers the title’s question, with two important qualifications attached. First, “runs on 8GB” is a claim about the INT8 build only, and BFL quantifies no quality loss from that quantization. Second, every quality ranking in the release is BFL’s own internal evaluation, so the buying decision should rest on the VRAM budget, the license terms, and your own prompt suite, not on the vendor’s leaderboard.
What shipped: five builds of one model
Klein is a single model distributed in several forms. The release covers text-to-image generation, image editing, and multi-reference generation with up to 10 reference images, per BFL’s announcement. The builds worth knowing:
- Base and distilled BF16: full-precision weights at 11.2 GB. The distilled variant trades generation steps for speed.
- FP8: 5.6 GB, the smallest weight footprint, though practical memory use includes activations and overhead on top of weights.
- INT8: the build BFL says fits 8GB RTX cards.
- NVFP4: a build aimed at NVIDIA’s newest datacenter hardware, where the company reports sub-second generation.
Day-one tooling support is unusually broad for an open image model: ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni, all listed in the release. LoRA fine-tuning is supported from day one in Diffusers and Ostris’s AI Toolkit, with NVIDIA NeMo AutoModel recipes shipping against the released checkpoints.
The VRAM ladder, mapped to real cards
The decision-relevant numbers are the weight footprints, because those set the floor for what each GPU tier can load:
| Build | Weight footprint | GPU tier it fits | Notes |
|---|---|---|---|
| BF16 | 11.2 GB | 12GB cards are tight; 16GB+ comfortable | Full precision, base or distilled |
| FP8 | 5.6 GB | 8GB cards, with headroom for activations uncertain | Smallest documented weights |
| INT8 | Unspecified, but sized to fit | 8GB RTX cards per BFL | No quantified quality-loss figure |
| NVFP4 | Unspecified | NVIDIA GB300-class datacenter hardware | Sub-second generation claimed |
Read this table with care. The 11.2 GB BF16 figure means a 12GB card technically loads the weights but leaves little room for activations, the VAE, and ComfyUI’s graph overhead. The FP8 figure of 5.6 GB is smaller than the INT8 target hardware, which seems counterintuitive; the practical difference is that BFL explicitly states that the INT8 path runs on 8GB RTX GPUs through ComfyUI or Diffusers, while making no equivalent statement about FP8 on that tier. If you are buying or allocating hardware, the documented claim is INT8 on 8GB. Anything else on an 8GB card is an experiment you run yourself.
The BFL release that names Qwen-Image 2.1 and Krea 2 publishes no VRAM guidance for either of them. At 7B and 12.9B parameters respectively, their BF16 footprints would exceed Klein’s. For the smaller rival, one community log offers a calibration: a Qwen-Image 2.1 int8 checkpoint measured at 17.3 GB on a 24GB card, per a ComfyUI 24GB-GPU repository. That is a single datapoint, not a spec, and the gap in vendor guidance is not evidence the rivals cannot fit smaller cards; it is evidence you cannot plan around them from this release.
Measured speed, with an asterisk on the comparison
BFL’s speed claims, measured by NVIDIA on the quickest build per GPU, according to the release: with the distilled model and INT8 build, Klein generates a 1MP image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 through ComfyUI, at 16 steps. On NVIDIA’s GB300 in NVFP4, the distilled model is reported at under 0.6 seconds per 1MP image, and serving a LoRA adds around 5% to generation time, per the release.
Those are usable planning numbers for single-GPU deployment. Under 9 seconds on a 4070-class card puts interactive editing within reach on consumer hardware, which is the actual news for teams that previously assumed local generation meant waiting.
The cross-model latency chart in the same release is less usable. BFL’s own methodology note states that “FLUX 3 Image Klein was timed on our servers. Other models were timed from the client, which includes polling and network time.” That asymmetry inflates every rival’s apparent latency, and it means the chart cannot support a “Klein is faster than Qwen-Image 2.1” conclusion. The same disclosure notes Krea 2 was timed on its Medium tier while its Large tier took 31.1 seconds for text-to-image, so tier selection alone moves that number substantially.
What Qwen-Image actually proves, and why the 2.1 evidence is thinner
The strongest documented reason to prefer Qwen-Image over Klein is text rendering, especially Chinese. The Qwen-Image technical report records a score of 58.30 overall on its ChineseWord benchmark, against 36.14 for GPT Image 1 [High] and 33.05 for Seedream 3.0. In the report’s qualitative comparisons, “only Qwen-Image and GPT Image 1 [High] manage to render such a long text clearly and almost perfectly” on a long-paragraph case, while other models omitted words or produced duplicates.
The report also describes a methodology worth emulating: each competitor in its arena, including FLUX.1 Kontext [Pro] and Ideogram 3.0, was entered into at least 10,000 pairwise comparisons by more than 200 invited evaluators. That is a scale and disclosure standard BFL’s internal Klein evaluation does not document.
Two cautions apply before you route a decision on this evidence. The report covers the original Qwen-Image, not Qwen-Image 2.1; whether 2.1 preserves the text-rendering edge is not established here. And ChineseWord is Qwen’s own benchmark, so the 58.30 figure carries the same vendor-origin caveat as BFL’s Elo numbers. The difference is that the Qwen report publishes a reproducible methodology, which gives you something to rerun on your own prompts.
If your workload is dominated by Chinese-language text in images, long embedded paragraphs, or typography-heavy output like posters and UI mockups, the original model’s documented strength makes Qwen-Image 2.1 worth testing even without an 8GB story. If your workload is photorealistic product shots or multi-reference editing, the Qwen-side evidence says much less.
How to read the 91-Elo claim
BFL’s headline quality claim is that Klein “ranks first in text-to-image, multi-reference generation and image editing, ahead of larger open weights models,” rating above the closed Ideogram V4.5 and FLUX.2 [pro]. Against FLUX.2 [klein] 9B, per the release, it “gains 91 Elo in text-to-image and 70 in image editing, with 38% fewer parameters.” For context, FLUX.2 [pro] supports up to 4MP output and 10-image multi-reference, so claiming parity above it at 5.6B parameters is an aggressive positioning statement.
All of that is internal evaluation. Research on Elo-based arena rankings in the LLM world gives a concrete reason for caution: Elo scores are “a noisy metric, because it is inherently subjective and can change from day-to-day,” and each score “comes with a 95% confidence interval” wide enough to shift rankings, per the model-consistency study. That work concerns chatbot arenas, not image models, so treat it as an analogy rather than a measurement of BFL’s chart. But the underlying math is the same, and BFL publishes no confidence intervals, rater counts, or prompt sets for its deltas.
A 91-point Elo gap is large enough that it probably reflects a real preference direction against FLUX.2 [klein] 9B, BFL’s own previous model, which is also the comparison where BFL has the least incentive to err. The claims against Qwen-Image 2.1 and Krea 2 deserve more discounting, because the evaluation conditions for those models are undescribed.
The practical alternative to trusting vendor Elo is running metrics you can reproduce locally. The Vendi Score paper catalogs the standard families for image generators: Inception Score, FID, Precision and Recall, plus diversity-oriented measures like the Vendi Score itself. None of these capture text-rendering accuracy, so pair them with a fixed set of your own prompts covering your actual use cases, scored by human reviewers on your team. That suite, run identically against Klein INT8 and Qwen-Image 2.1, is worth more than either vendor’s chart.
License and cost math: where the $5M line sits
The FLUX Kommunity License “allows free commercial use for individuals and companies with annual revenue under $5M, alongside research and non-commercial use,” per BFL’s release. Above that line, commercial deployment requires BFL’s self-hosted licensing or its API, and the release publishes no self-hosted license prices.
Two compliance details matter. The threshold is total annual company revenue, not revenue from your image product, so a large enterprise running Klein in a side tool still crosses it. And the threshold is presumably assessed at the time of use, so a startup approaching it should price the transition before it arrives rather than after.
Against self-hosting, BFL prices the flux-3-image-klein API, per the release, starting at $0.009 per 1K image for the distilled model with prompt upsampling and pixel restoration. The break-even arithmetic favors the API at low volumes almost regardless of hardware cost: a used RTX 4070-class card plus its power and your maintenance time takes a long while to pay back at fractions of a cent per image. Self-hosting wins on different grounds: data sovereignty, where images cannot leave your hardware; latency guarantees under your control; on-device editing with no network dependency; and freedom from per-image metering at volumes where per-image fees add up. It also wins if you need fine-tuned weights in production, since the day-one LoRA path lets you serve adapted models BFL does not host.
One unresolved cost sits above that line: BFL’s self-hosted license pricing is not published, so any total-cost model for a larger organization has an unknown term in it. Request a quote before committing an architecture to Klein.
Fine-tuning is the real bottleneck
The release’s most durable consequence may be the shift of the constraint from weights to data. An 8GB floor means the model itself is no longer the scarce resource for most teams; a workstation card suffices. What remains scarce is the fine-tuning pipeline: a curated dataset for your domain, the LoRA training runs in Diffusers or Ostris’s AI Toolkit, and the evaluation loop to confirm the adapter helps.
BFL reports that serving a LoRA adds only about 5% to generation time, so the serving cost of customization is small. The cost sits upstream, in collecting and cleaning examples that capture your house style, product line, or editing conventions. Budget for that work, and for the license review, before budgeting for GPUs. Teams that skip the data investment get a general-purpose model that any competitor can download the same day.
The decision, stated plainly
On the evidence available, the 8GB answer is FLUX 3 Image Klein’s INT8 build through ComfyUI or Diffusers, because BFL’s release documents it as an 8GB fit while publishing no comparable guidance for Qwen-Image 2.1 or Krea 2. Choose it for sovereignty-constrained pipelines, on-device editing, and fine-tuned production workloads, provided your organization’s annual revenue is under the license’s revenue threshold or you have priced BFL’s self-hosted license. Expect roughly 4 to 9 seconds per 1MP image on RTX 5090 to 4070-class cards under the best-case conditions BFL and NVIDIA measured, and verify on your own hardware.
Choose Qwen-Image 2.1 when Chinese or long-text rendering dominates the workload and you have headroom above 8GB, but recognize that its documented strengths come from the original Qwen-Image report, and that a community measurement of its int8 build, 17.3 GB, points at a substantially larger card. Treat every ranking in the Klein release as vendor-directional until independent benchmarks appear, which, given the two-day-old release, they likely will within weeks.
Before shipping either model, run a fixed prompt suite from your real workload against both, score it yourself, and spot-check with FID or Precision/Recall if you need distributional numbers. The vendor charts tell you where to start testing. They cannot tell you what to ship.
Frequently Asked Questions
Does FLUX 3 Image Klein run on an 8GB GPU?
With the INT8 build, FLUX 3 Image Klein runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM, through ComfyUI or Diffusers.
What is the revenue limit for free commercial use of FLUX 3 Image Klein?
The FLUX Kommunity License “allows free commercial use for individuals and companies with annual revenue under $5M, alongside research and non-commercial use,” per BFL’s release.
Join the discussion
Share a useful perspective or ask a question about this article.