Open Weight Model Licenses: Llama vs Mistral vs Qwen IP Clauses for SaaS Build or License

This guide compares the IP clauses of Llama, Mistral, and Qwen open-weight licenses for SaaS teams deciding whether to build or license in 2026.

It walks through a three-gate check covering commercial hosting rights, downstream use-restriction flow-down, and EU AI Act documentation duties under Article 53(1)(b) before the August 2, 2026 enforcement date.

How the license-to-SaaS chain actually works

The license-to-SaaS chain starts with a definitional move that most teams miss. Under the EU AI Act, a downstream provider is anyone who provides an AI system that integrates a general-purpose AI model — which is exactly what a SaaS product does when it wraps Llama, Mistral, or Qwen behind an API or a chat surface. That status is not optional and not something you can contract away with the model vendor. It also carries an entitlement: under Article 53(1)(b) and Annex XII, the downstream provider is entitled to documentation from the model provider (mmoww.net). Read that as a gate, not a courtesy. If you cannot obtain the model documentation you are entitled to, you cannot demonstrate the integration you are shipping.

The second link in the chain is the downstream obligation itself. As defined in the NHI Management Group glossary, a downstream obligation is the continuing responsibility to honour a security, privacy, or access requirement after data, credentials, or records move beyond the original controller's environment (nhimg.org). The key word is continuing. The first system does not end the duty. In practice this is where SaaS teams get surprised: personal data flows into your inference layer, prompts and outputs land in logs, service accounts and API keys stay active in external systems, and the deletion, correction, restriction, and access duties follow all of it. The most common misapplication, per the same source, is assuming a processor or tool vendor automatically inherited the same deletion or access-response duty when only the data moved and not the accountability model.

The third link is flow-down, and this is the mechanism that turns a model license into your problem. Downstream flow-down is the practice of requiring subcontractors and lower-tier providers to inherit the same security, privacy, and reporting obligations that apply to the prime supplier (nhimg.org). Applied to a SaaS stack, it means the obligations you carry as a downstream provider must be pushed to every subprocessor, hosting provider, and fine-tuning vendor beneath you. The license you signed with the model provider is the upstream end of that chain; your own customer contracts are the downstream end. If the two ends do not match, the gap is yours.

Run the chain as a three-gate check before any EU ship. Gate one: does the model license permit commercial SaaS hosting at all? Gate two: does it require downstream flow-down of use restrictions, and to whom? Gate three: does it obligate the model provider to hand you the Article 53(1)(b) and Annex XII documentation you are entitled to? A license that clears gate one but fails gate two or three leaves you holding an obligation you cannot pass along and cannot evidence.

Treat the license as the upstream anchor and your customer contracts as the downstream anchor, then verify the chain holds at every hop. The clause that decides your exposure is the one that determines whether the obligation travels with the model — not the one that determines whether you can host it.

The evidence: dates, fines, and documentation

The dates are the part of this analysis that teams get wrong most often, and they are the part that carries the money. EU AI Act GPAI model obligations went live on August 2, 2025, according to risktemplate.com, which means the documentation duties attached to general-purpose models are already in force rather than pending. The enforcement mechanism that gives those duties teeth arrives on August 2, 2026, when fines of up to 3% of global turnover become available against providers who are not compliant. Read those two dates as a single window: the obligation is live now, the penalty is live soon, and any SaaS product shipping into the EU inside that window is exposed on both sides of it.

The specific duty that reaches a SaaS team is documentation, not permission. Article 53(1)(b) requires model providers to supply downstream providers with documentation under Annex XII, as mmoww.net sets out. That matters because a downstream provider under the Act is anyone who provides an AI system that integrates a general-purpose AI model — the definitional position a SaaS product occupies the moment it wraps a model behind an API. The practical check is to confirm, in writing, that your model supplier has delivered Annex XII documentation and that your own downstream customers can receive the equivalent from you. If that documentation does not exist in your vendor file, the August 2026 exposure is yours to carry, not your model provider's.

The third gate is where the most common misapplication sits. As nhimg.org explains, the frequent error is assuming a processor, partner, or tool vendor has automatically inherited the same deletion or access-response duty when only the data has moved, but not the accountability model. Moving data does not move responsibility by itself. So when a SaaS team routes prompts or embeddings through a third-party tool, the check is not whether the data arrived — it is whether the receiving party is contractually bound to the same deletion, correction, restriction, and access duties your own customers can invoke. If that flow-down language is absent, the obligation stays with you.

Run the three gates as a sequence, not a checklist. Gate one fails and the license question is closed regardless of documentation quality. Gate two fails and you need flow-down language in your customer terms before you can rely on the license at all. Gate three fails and you are carrying an Article 53(1)(b) documentation gap into an enforcement date that is measured in months, not years. The license you pick among Llama, Mistral, and Qwen determines which gates are easy; the August 2, 2026 date determines which gates you cannot skip.

Llama vs Mistral vs Qwen: clause-by-clause

The three licenses diverge on exactly the three clauses your three-gate check turns on: commercial SaaS hosting, downstream flow-down, and documentation. Start with Llama. The Meta Community License is not open source in the OSI sense, and its two operative clauses for a SaaS team are the acceptable-use policy and the monthly-active-user threshold that triggers a separate license grant. The available sources for this analysis does not state that threshold, so do not rely on a remembered number: open the license URL, search the text for the MAU clause, and confirm the figure and the trigger condition before you ship. If your product crosses it, the license you thought you had is no longer the license you have.

Mistral is the license most likely to be misread, because the answer depends on which license attaches to the specific model version you pulled. Some Mistral releases ship under Apache 2.0, which permits unrestricted commercial SaaS hosting with no flow-down and no user threshold. Others ship under the Mistral Research License, which does not. There is no shortcut here: go to the Mistral license page for the exact model version in your dependency manifest, confirm which of the two applies, and record the version string next to the license name. A team that checked Mistral once and generalized to "Mistral is Apache" has an unverified assumption, not a compliance position.

Qwen requires the same treatment on a different axis. The Tongyi Qianwen License carries a commercial-use trigger and its own downstream restrictions, and both need to be read against your deployment shape rather than against the model card summary. Check the trigger condition first — it determines whether commercial use is permitted at all — then check whether the license imposes obligations on parties further down your stack. If it does, those obligations are the ones your own terms of service must pass through to your customers.

Read the table as a checklist, not a verdict. Each row's middle two columns are the gates; the right column is the only place the answer is authoritative. The comparison this section makes is narrow on purpose: on commercial SaaS, flow-down, and documentation, the three licenses are not interchangeable, and the differences sit in clauses you can read in an afternoon. What the licenses do not tell you is how those clauses interact with the EU AI Act's downstream-provider obligations — that interaction is where the enforcement exposure actually lives, and it is treated separately.

Costs and numbers that matter

The penalty exposure priced above — up to 3% of global turnover from August 2, 2026, per risktemplate.com — is the number that should anchor the whole decision, and it is not a per-model fine. It is calculated against the SaaS provider's worldwide revenue, not against the value of the model contract or the size of the EU deployment. A team running a modest EU pilot on a product with nine-figure global revenue should treat the ceiling as a board-level number, because the exposure scales with the company, not the workload.

Documentation is the second cost line, and it is the one most likely to be underestimated. Article 53(1)(b) and Annex XII entitle a downstream provider to obtain specified documentation from the model provider, per mmoww.net. The operative word is "obtain." If the license you accepted does not contractually guarantee delivery of that documentation, the obligation to produce it does not disappear — it converts into legal review hours spent reconstructing training-data summaries, intended-purpose statements, and technical documentation from public model cards, release notes, and your own integration records. Before signing, check whether the license names Article 53(1)(b) and Annex XII deliverables explicitly; if it does not, budget the reconstruction as a line item rather than a contingency.

Flow-down is the third cost driver, and it compounds per sub-processor: as defined above, every downstream clause in the model license adds a contract-review checkpoint for each SaaS sub-processor that touches the model or its outputs. The arithmetic is unforgiving: a license with three flow-down clauses across six sub-processors generates eighteen review points, and each one is a place where a missing clause becomes an unenforced obligation.

The three-gate check from the reader rule maps directly onto these numbers. Gate one — commercial SaaS hosting — determines whether you owe the 3% exposure at all. Gate two — downstream flow-down of use restrictions — determines how many contract-review checkpoints you staff. Gate three — documentation guarantees — determines whether you pay in license negotiation or in legal reconstruction. Run all three before shipping into the EU, and price the worst case at 3% of global turnover, not at the model's list price.

What the evidence does NOT establish

Every analysis has a boundary, and this one has a hard one: the available sources behind the three-license comparison does not state several details you would need before signing anything. This section exists so you know exactly what remains unverified and where to go check it yourself.

First, the license-specific terms are not confirmed in the available sources. It does not state whether Llama's community license carries a 700 million monthly active user threshold, how Mistral's licensing splits across individual model versions, or what trigger Qwen sets for commercial use. Each of these could change your three-gate answer for a given model, so verify the current text at the license URL published by each vendor before you ship. License terms are also version-specific — the clause that governs one model release may not govern the next.

Second, the available sources does not establish whether any of the three licenses expressly require downstream flow-down of the Article 53(1)(b) documentation duties. The mmoww.net source confirms that downstream providers are entitled to that documentation under the Act itself, but entitlement under the regulation is a separate question from whether the license contract compels the model provider to hand it over. Check the license text for explicit flow-down language — per nhimg.org, obligations often fail at handoff, not at creation, and the most common misapplication is assuming accountability automatically transferred when only the data or weights moved.

Third, the available sources does not state whether Annex XII documentation is actually available, per model, from any of the three vendors. The regulation gives you the right to it; it does not guarantee the vendor has produced it in usable form. Request it from each model provider in writing before you commit, and treat a slow or evasive response as a finding in itself.

One more caveat on timing: the risktemplate.com source puts enforcement fines of up to 3% of global turnover at August 2, 2026, but enforcement posture in individual member states is not addressed by the available sources. Confirm how your lead EU market is transposing and enforcing before you finalize the compliance plan.

EU SaaS on an open-weight model

Let's run a concrete scenario through the three-gate check, against the enforcement date and penalty ceiling already priced above. Your team is building a SaaS product that hosts an open-weight model, you serve EU customers, you work with three sub-processors (an inference host, a logging vendor, and a support-ticket system), and you want to launch before the enforcement date. Here is the gate-by-gate walkthrough.

Gate 1: Commercial SaaS permitted? Go to the actual license URL for the model you picked — not a blog summary of it — and search for the hosting and commercial-use terms. If the license is silent on commercial SaaS hosting or restricts it, you stop here, full stop. A license that permits you to run the model internally may not permit you to sell access to its outputs, and if Gate 1 fails, no amount of sub-processor work downstream will fix it. Silence counts as a restriction for your purposes; get a written legal opinion or move on.

Gate 2: Flow-down required? Count the downstream clauses in the license — the ones requiring your customers or providers to inherit use restrictions. Each one you find becomes a contract amendment obligation for your sub-processors, applying the flow-down mechanism defined above. In this scenario, three sub-processors means three amendments. If the license carries two flow-down clauses per contract, you are drafting six clause additions across three vendor agreements. Count them before launch, because obligations often fail at handoff, not at creation.

Gate 3: Documentation available? As a downstream provider under the EU AI Act, you hold the documentation entitlement described above under Article 53(1)(b) and Annex XII. Check before launch whether the model publisher actually supplies that documentation package. If it does not, your compliance file has a hole you cannot patch on your own.

The practical sequencing: run Gate 1 first, because it is binary; run Gate 2 second, because its cost scales with your sub-processor count; run Gate 3 third, as a documentation audit. A SaaS team that passes all three gates before August 2, 2026 ships with an enforceable chain; a team that skips Gate 2 ships with amendments still in redline.

Decision rules for the license register

The register is where the three-gate check stops being a checklist and becomes a set of if/then rules you can actually run. Write them as decision rules, not as notes, because the point of the register is to force a binary outcome at each gate before a model is selected. The first rule is the one that carries the most weight: if the SaaS serves EU customers and the launch date falls after August 2, 2026, then require Article 53(1)(b) documentation in writing before selecting the model. That documentation right is the downstream provider's entitlement under the EU AI Act, and per mmoww.net it covers the model provider's obligations under Article 53(1)(b), read with Annex XII.

Annex XII. Do not accept a link to a model card as a substitute; the register entry should record the specific document received, the date received, and the name of the person who confirmed it.

The second rule governs contract structure. If the license contains more than two downstream flow-down clauses, then add a sub-processor contract amendment step to the register. Downstream flow-down, per nhimg.org, is the practice of requiring subcontractors and lower-tier providers to inherit the same security, privacy, and reporting obligations that apply to the prime supplier. Two clauses is the threshold because a pair can usually be handled inside the standard terms; a third clause almost always means the obligation has to travel to your own sub-processors, and that travel requires a contract amendment rather than a checkbox. The register entry should name each sub-processor affected and the amendment status for each.

The third rule is the escalation rule. If the model provider will not confirm Annex XII documentation, then treat the license as failing Gate 3 and escalate to counsel. This is not a negotiation position; it is a stop. The register should record the refusal, the date of the request, and the counsel escalation, so that the decision to proceed or not is documented rather than assumed. A provider that will not confirm the documentation has effectively removed the downstream provider's statutory footing, and the register needs to show that the team recognized it.

Run the three rules in order and the register produces one of three states per model: cleared, cleared with amendment, or escalated. The table below is the register format I use, with the rule that triggers each state.

Two operational notes keep the register honest. First, the rules are cumulative: a model can clear Gate 1 and still fail Gate 3, and the register should never collapse the three states into a single pass/fail. Second, the register is a pre-selection artifact, not a post-launch audit trail. If a rule fires after the model is already in production, the register entry should say so explicitly, because the remediation path differs from the path for a model that has not yet shipped.

What to do next

Frequently Asked Questions

When do the EU AI Act's GPAI enforcement fines actually start applying?

Enforcement with fines of up to 3% of global turnover begins August 2, 2026, after the GPAI model obligations went live on August 2, 2025.

What documentation is a downstream provider entitled to receive from a GPAI model provider?

Under Article 53(1)(b) read with Annex XII, a downstream provider integrating a GPAI model is entitled to documentation from the model provider.

Who counts as a downstream provider under the EU AI Act?

A downstream provider is anyone providing an AI system that integrates a general-purpose AI model.

What is the first gate in the three-gate check for SaaS teams?

Gate 1 requires verifying whether the model license (Llama community license, Mistral Apache-style, or Qwen terms) expressly allows offering the model as a hosted commercial service.

What is the second gate in the three-gate check?

Gate 2 requires verifying whether the license requires downstream flow-down of use restrictions.

Which IP clause is decisive for EU-bound SaaS teams comparing these licenses?

The decisive IP clause is the downstream obligation flow-down required by the EU AI Act's August 2, 2026 enforcement date, not the model license text alone.

Quick answers

Also worth reading: 1 in 5 SaaS Licenses Cite Dead Patents: The 20-Year Clock: 1 in 5 SaaS Licenses · API Triangulation and Clearance Velocity Data for IP Teams: API Triangulation and Clearance Velocity