← Journal

Black Forest Labs caps free commercial use at $5M

Black Forest Labs opened FLUX 3 Image Klein only to companies under $5 million in revenue, while Meta's open 816B model needs two 512GB Mac Studios.

Three model releases landed on the same day, and the part that decides what each one is worth is not the parameter count. It is the sentence naming who may use it.

**Black Forest Labs** opened the weights of FLUX 3 Image Klein, a 5.6 billion parameter image model with native 4K output and bounding box placement, under a licence free for commercial use only by companies below $5 million in annual revenue, VentureBeat reported. The hardware bar is deliberately low: the INT8 build runs on an Nvidia GPU with 8GB of VRAM and generates an image in under 9 seconds on an RTX 4070. The model is open to anyone with a gaming card and closed to anyone with a business above the line.

Against the ledger's record that is continuity, not a turn. The Klein line has been the on-device tier since Black Forest Labs shipped FLUX.2 Klein on ASUS ProArt laptops in June (Black Forest Labs), and its CEO Robin Rombach urged G7 leaders to back open innovation two weeks after that (Black Forest Labs). The company also raised $300M at a $3.25B valuation in December 2025 (Crunchbase News), which is what makes a revenue ceiling legible: the weights are the distribution, and the licence is where the revenue stays.

Three ceilings, in three different places

Meta released the weights of Muse Spark Open, an 816B parameter mixture of experts model with 42B active parameters and a 1M token context window, on Hugging Face under the Muse Model License, and says it runs on two 512GB Apple M5 Ultra Mac Studios (TechCrunch). No revenue test there. The ceiling is the machine, and behind it a second one: Meta says the open model is less capable than the closed Muse Wonder it showed earlier in the week.

Microsoft put its ceiling in the contract. It said it finished pre-training MAI-Base-2, a natively multimodal mixture of experts model with 2.8 trillion total parameters and 80 billion active per token, trained on 50 trillion tokens across 16,000 Nvidia GB200 GPUs over 38 days, and that the weights will be licensed to partners through Foundry with no release date named (GIGAZINE).

So, in one day: a revenue test, a hardware bill, and a partner list. Each of the three gets described as open by someone, and no two of them are the same promise.

Why the ceiling is where the argument is

The same week, the 20 most valuable global AI and semiconductor stocks shed more than $1 trillion in market value, with SK Hynix down more than 10% and Samsung Electronics about 8.7%, Dell down 6.8% and the Philadelphia Semiconductor Index about 23% below its June peak, as a strong US inflation report revived concern about the cost of financing AI buildouts (CNBC). Microsoft's figure states that concern in hardware: 16,000 GB200s, 38 days, for a model nobody outside its partner list will hold.

Which is the reason a licence clause does the work a price used to do. Giving away weights costs nothing per copy and recovers nothing per copy, so the recovery has to be written into the terms:

  • Black Forest Labs recovers above $5 million in customer revenue.
  • Meta recovers from anyone who cannot assemble the hardware, and keeps its better model closed.
  • Microsoft recovers through Foundry, from partners it names.

Meta showed the same instrument pointed the other way on the same day. Its updated commercial terms, effective October 30, authorise it to crawl websites associated with business accounts on its platforms and use that content to train its AI products with no opt-out, and legal experts told Nieman Lab that publishers who keep using Meta's business tools after that date would effectively have consented. The download is free either way. The terms decide what it is worth, and who it is worth it to.

Built from the digest of 2026-10-16.