TL;DR: “Open weights” and “free for commercial use” are different claims, and several of the most-downloaded models fail the second one. FLUX.1-dev, Mistral Large, and Command R+ prohibit commercial use outright; Stable Diffusion 3.5 cuts off at $1M revenue; Llama 4 caps you at 700M monthly users. If you charge money for anything the model touches, check the license before you build.

	Non-commercial license	Revenue-capped	MAU-capped	Truly permissive
Examples	FLUX.1-dev, Mistral Large (MRL), Command R+	Stable Diffusion 3.5	Llama 4, Qwen large tiers	Qwen3.6-35B-A3B, DeepSeek V4, FLUX.1-schnell
Commercial self-hosting	❌ Requires paid license	✅ Under $1M/yr revenue	✅ Under user cap	✅ Unrestricted
Cost to go commercial	Negotiated fee or royalty	Custom Enterprise License	Negotiated with vendor	$0
The catch	”Non-commercial” includes your day job	Threshold counts ALL revenue	Plus attribution and regional clauses	None — this is the bar

Honest take: Never build a product on a model whose license page says “contact sales” anywhere. In every category listed below, an Apache 2.0 or MIT alternative exists within 10% of the quality — take the clean one and skip the legal review.

Model release posts in 2026 lead with benchmark deltas and bury the license in a footnote. That footnote is the difference between shipping a product and receiving a letter from a vendor’s legal team. Below is the current blacklist — the widely used “open” weights that block or cap commercial use — with the exact clause that does the blocking, verified against the license texts in September 2026, and the clean-licensed swap for each one.

Which “open” model licenses ban commercial use outright?

Three headline releases are non-commercial, full stop: FLUX.1-dev, everything under the Mistral Research License, and Cohere’s Command R family. Downloading is free; charging money for what they produce is not licensed.

FLUX.1-dev (Black Forest Labs) ships under the FLUX.1 [dev] Non-Commercial License. You may use, distribute, and fine-tune the model for non-commercial purposes only; anything else requires a separate license from BFL, granted “in Company’s sole discretion” and potentially subject to “a fee, royalty or other revenue share.” The trap most people miss: early license text explicitly allowed commercial use of generated outputs, but the v1.1 revision (June 2025) removed that line and now defines Non-Commercial Purpose to exclude any use where you receive direct or indirect payment arising from the outputs. Selling images made with a self-hosted FLUX.1-dev is no longer covered. FLUX.1-schnell — same family, Apache 2.0 — has no such restriction.

Mistral Large and every model marked MRL use the Mistral AI Research License, which permits only “personal, scientific or academic research” and non-profit use. Two clauses make it stricter than it sounds. First, it excludes use “by individuals or contractors employed in or engaged by companies in the context of their daily tasks” — so running Mistral Large as your work coding assistant is already outside the license, even if you never ship it to a customer. Second, it excludes “any testing or proof-of-concept that is intended to generate revenue” — you cannot even legally prototype a commercial product on MRL weights. Commercial self-deployment requires a negotiated Mistral Commercial License. Mistral’s Apache 2.0 releases (the small and Nemo lines) are unaffected; check the license field per checkpoint, not per company.

Command R and Command R+ (Cohere) are released under CC-BY-NC 4.0 plus Cohere’s acceptable-use addendum. Creative Commons NonCommercial means what it says: research, students, and hobbyists are fine; any commercial use of the weights needs a deal with Cohere’s sales team. Cohere’s later Command A went Apache 2.0, so even within one vendor the answer flips model by model.

What counts as “commercial use” for a self-hoster?

More than most people assume: internal business tooling, client work, and monetized side projects all count — not just selling API access. The three licenses above draw the line at whether the use is connected to commercial activity, not whether you resell the model itself. Concretely, under a non-commercial license you are outside the terms if you:

• Run the model as an internal assistant at a company (MRL names this case explicitly)
• Use its outputs in paid client work — logos, copy, code
• Put its outputs on a monetized blog, YouTube channel, or ad-supported site (“indirect payment” in FLUX.1-dev v1.1 language)
• Prototype a startup idea on it, even pre-revenue, if the intent is commercial

What stays inside the terms: a homelab you don’t monetize, academic research, personal projects, and evaluating models to decide what to buy. If your setup is genuinely hobbyist, none of this article applies to you yet — the danger is that side projects have a habit of becoming products, with the license unread the whole way.

Which licenses allow commercial use but cap it?

The second tier is “commercial until you succeed”: Stability AI caps by revenue, Meta and Alibaba cap by user count. These are usable in production, but the cap is a clause your future acquirer’s lawyers will find.

Stable Diffusion 3.5 (and SD3) use the Stability AI Community License: free commercial use for individuals and organizations under $1M in annual revenue — total revenue from any source, not revenue attributable to the model. Cross the threshold and you must buy a custom-priced Enterprise License to keep using the weights commercially. For a solo developer this is generous; for an established business it is a mandatory purchase disguised as an open release.

Llama 4 uses the Llama Community License: commercial use is allowed until you exceed 700 million monthly active users, at which point you must request a license Meta may or may not grant. It also requires “Built with Llama” attribution, requires fine-tune names to start with “Llama,” and — the clause that actually bites in practice — Meta’s Llama 4 terms exclude EU-domiciled companies from multimodal use entirely. The MAU cap will never bind a homelab; the EU exclusion and attribution rules get Llama-based stacks rejected in enterprise procurement today. Our open-source LLM licensing guide walks through the full text field by field.

Qwen’s largest tiers are the quiet one. Alibaba releases most Qwen models at or below 32B under plain Apache 2.0 — that’s why Qwen3.6-35B-A3B tops our license shootout. But the biggest checkpoints have historically shipped under the separate “Qwen License,” which adds a 100M-MAU threshold requiring a commercial deal with Alibaba. Same family, same model card layout, different legal reality — verify the license file on the exact checkpoint you pull, not the family’s reputation.

Gemma sits at the edge of this tier: Google’s Gemma Terms of Use permit commercial use with no revenue or user cap, but bind you to a prohibited-use policy that Google can update after you deploy. It is not OSI open source, and a moving use-policy is itself a string — a mild one, but worth knowing it’s there.

How do you check a license before you build on it?

Read the actual LICENSE file on the exact checkpoint, then search it for five phrases. Takes two minutes, beats any headline:

# Pull just the license file, not 40GB of weights
curl -sL https://raw.githubusercontent.com/black-forest-labs/flux/main/model_licenses/LICENSE-FLUX1-dev \
  -o LICENSE.txt

grep -inE "non-commercial|monthly active|annual revenue|sole discretion|research purposes" LICENSE.txt

Expected output on a restricted model (FLUX.1-dev shown):

2:FLUX.1 [dev] Non-Commercial License
...multiple hits on "Non-Commercial Purposes" and "sole discretion"...

On a genuinely permissive model the grep comes back with zero hits, because MIT and Apache 2.0 contain none of those phrases. Any hit means stop and read the full text. Two more signals worth 30 seconds each: a Hugging Face repo that makes you click through a license acceptance gate is never plain Apache/MIT, and a license hosted on the vendor’s own domain (rather than a standard SPDX identifier in the metadata) usually means custom terms.

One September 2026 caveat: license files change in place, as FLUX.1-dev’s v1.1 revision showed. If you’re betting a product on the terms, save a dated copy of the license text you accepted.

What does switching to a clean-licensed model cost?

Usually nothing in hardware, because the permissive alternative in each category fits the same VRAM tier. Prices below verified September 2026, from the comparisons above:

Your situation	License-safe pick	Hardware floor	Cost	Where
Image generation, was eyeing FLUX.1-dev	FLUX.1-schnell (Apache 2.0)	12-16GB VRAM, same as dev at Q8 GGUF	$0 swap on existing GPU	Already have it
LLM under 24GB, was eyeing Mistral Large or Command R+	Qwen3.6-35B-A3B (Apache 2.0), ~21GB at Q4_K_M	Used RTX 3090 24GB, ~$1,070 used (Sep 2026)	~$1,070 once	Check price
Want to validate output quality before committing	Rent the same GPU hourly	none owned	RTX 3090 from $0.07/hr (Sep 2026, market rate)	Vast.ai

The quality gap argument for accepting a restricted license has mostly evaporated. In text, MIT-licensed DeepSeek V4 and Apache Qwen3.6 sit at or above Mistral Large’s tier; in images, FLUX.1-schnell at Q8 is close enough to dev that a license fee negotiated “in sole discretion” cannot be the rational choice for a small operation. GPU sizing for whichever model you land on is covered in depth at runaihome.com.

When NOT to worry about any of this

If nothing in your stack earns money and nothing is deployed for an employer, skip the audit — every license in this article permits personal, research, and evaluation use. Specifically, don’t burn time on license review when:

• You self-host purely for private use and have no monetization anywhere in the pipeline
• You’re benchmarking candidates before a decision — evaluation is fine under all of these terms
• You’re in an academic context, which MRL and CC-BY-NC explicitly welcome

The reverse also holds: don’t treat this article as legal advice for a funded company. Past roughly $1M revenue or any enterprise procurement process, the license question moves from “read the file” to “counsel signs off on the stack,” and the cost of that review belongs in your true year-one self-hosting budget alongside the hardware.

Verdict

Treat “open weights” as a download instruction, not a usage right. As of September 2026 the blacklist for commercial self-hosters is short and stable: FLUX.1-dev, anything marked MRL or MNPL (Mistral Large, Codestral’s non-production builds), and Command R/R+ are non-commercial without a paid deal; SD 3.5 stops being free at $1M revenue; Llama 4 and Qwen’s largest tiers carry MAU caps plus attribution or regional clauses. Every one of them has an Apache 2.0 or MIT neighbor at comparable quality — Qwen3.6-35B-A3B, DeepSeek V4, FLUX.1-schnell, Command A — so for a commercial build the decision is one grep and then the clean model, every time.

FAQ

Can I sell images I generated with a self-hosted FLUX.1-dev? Not under the current license. The v1.1 text (June 2025) defines Non-Commercial Purpose to exclude any use where you receive direct or indirect payment arising from the model’s outputs, and earlier language explicitly permitting commercial outputs was removed. Use FLUX.1-schnell (Apache 2.0) for anything sold.

Does a non-commercial license apply if I only use the model internally and never expose it to customers? Yes. Internal business use is commercial use under these licenses — the Mistral Research License spells out that employees using a model “in the context of their daily tasks” are outside research terms. Revenue exposure, not customer exposure, is the test.

Are fine-tunes of a restricted model any freer than the base weights? No. Derivatives inherit the base license in all the licenses covered here — an MRL fine-tune is still MRL, a Llama fine-tune keeps the Community License terms (and must keep “Llama” in its name). Only a model trained from scratch escapes the parent’s terms.

Sources

• FLUX [dev] Non-Commercial License — Black Forest Labs
• Mistral AI Research License (MRL)
• Stability AI Community License announcement
• Cohere Command R+ model card and license discussion — Hugging Face
• FLUX.1 model licenses — black-forest-labs/flux on GitHub

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe