Products

FLUX 3 Goes Closed-Source: A Heavy Blow to Open-Weight Image Models

Black Forest Labs pulled the open-weight card it played with FLUX 1 and now serves FLUX 3 only via API on its platform. The move narrows the runway for Chinese image models.

Font size

FLUX 3 no longer ships the weights

FLUX.1 launched under Apache 2.0, and anyone running ComfyUI locally knew it well. FLUX Pro straddled open and closed. With FLUX 3 Image, Black Forest Labs has settled the question: API only.

The news made it to the front page of Hacker News, hit 100 points, and gathered 11 comments — most of which missed the point. This isn't about the model getting stronger. It's about the business model changing tracks.

I read through the discussion. Everyone's still talking about prompt fidelity and text rendering — the usual technical angles. Nobody directly answered the question on my mind: when the team behind the strongest open-weight image model decides to sell only API access, what remains of the open-weight image model track?

Source: bfl.ai/models/flux-3-image

A generational jump, but where exactly

The official page highlights four capability buckets:

  • Photorealism
  • Prompt fidelity
  • Text rendering
  • Image editing

From the demo images, FLUX 3 improves over FLUX 1.1 Pro in fingers, fabric folds, and product-grade commercial imagery — but the gap with Midjourney v7 isn't a qualitative leap.

Some specific observations:

  • Fingers: six-finger and missing-finger artifacts were routine in the FLUX 1 era. The FLUX 3 demos show no obvious issues.
  • Fabric folds: shirts and silk finally show physically grounded subsurface scattering instead of that plastic look.
  • Commercial product shots: white-background product edges and shadow casting hit a commercially usable, straight-out-of-the-model level.
  • Text rendering: English typography on posters mostly lands correctly. Chinese remains a weak spot.

These improvements are catchable. FLUX 2, had it stayed open, could have closed the gap. What really separates the two paths is iteration speed — closed means no merge queues, no community PRs. Ship when it's ready.

Against the closed-source rivals, and the old open-source teammates

A side-by-side comparison:

DimensionFLUX 3 ImageMidjourney v7Ideogram 3.0DALL·E 4
PhotorealismStrongStrong+MediumGood
Prompt adherenceStrong+GoodGoodStrong
Text renderingGood (English)GoodStrongStrong
Editing capabilityStrongMediumMediumGood
Pricingbfl.ai usage-basedSubscriptionSubscriptionCredit-based

Within the closed-source tier, FLUX 3 still trails Midjourney by 5-10% on photorealism, but may overtake it on prompt obedience.

Against open-weight counterparts — Stable Diffusion 3.5, HunyuanImage, Qwen-Image:

  • Leads on technical benchmarks, but the ecosystem isn't in the same league.
  • ControlNet, LoRA, community workflows — the accumulated tooling of the SD lineage over many years. Nobody picks up the baton once FLUX closes up.

I saw people tuning FLUX LoRAs in ComfyUI not long ago. Those workflows vanish overnight once FLUX 3 closes its weights.

Where Chinese image models get squeezed: pricing power

Once FLUX 3 closes, bfl.ai sets the price and locks in the customer. When Chinese vendors pitch enterprise clients, the first question is, "why not just use FLUX?"

Open weights were once the moat for Chinese teams — if you won't give us an API, we'll run your weights ourselves, and the ecosystem position stays stable. Black Forest Labs just demolished that moat itself.

Specific landing scenarios:

  • E-commerce product shots: FLUX 3 ships commercially viable output directly. Chinese models need serious compute to catch up.
  • Poster design: text rendering plus layout — FLUX 3 is basically usable for English.
  • Content creation: subscription versus token-based billing — FLUX's model fits SMEs better.

Chinese models still have one path: go vertical. Medical imaging, industrial design, anime IP — long-tail scenarios where general-purpose competition is locked down by FLUX 3 and Midjourney v7.

Editorial notes

The FLUX lineage's drift mirrors a broader shift in the image model industry.

My read: closed-source earns the premium, open-weight is left playing catch-up. The weight of this matters because Black Forest Labs came up through the open-source camp. They proved open could produce state-of-the-art, then closed the door. That hits harder than OpenAI's nonprofit-to-LLC pivot or Meta's continued Llama releases.

What makes this different from past moves:

  • Unlike OpenAI's shift to closed, FLUX won SOTA while open — entirely different credibility.
  • Unlike Stable Diffusion's commitment to open, FLUX 3 demonstrates that ecosystem position can be eaten by pricing power.
  • Unlike Llama's "pseudo-open," FLUX 1's Apache 2.0 was genuinely open — which makes the reversal sting more.

Forecast: in the next 6-12 months, more image model teams will follow this path. Not because closed is inherently better, but because capital markets have already priced closed — bfl.ai can likely sustain the team on API revenue alone, even without a fresh round. Open-weight camps (SD lineage, Chinese teams) either find vertical moats or accept the "follower" label.

What I'm unsure about: whether FLUX 3 staying closed is sustainable depends on whether bfl.ai can absorb training costs. A single SOTA image model training run runs into eight-figure USD territory. Whether API revenue covers it — no public data. If it can't, the open route might come back in fashion. But that's a different story.

Source: https://bfl.ai/models/flux-3-image