Key Takeaways
- None of the ten model families ships under a single uniform license — terms are set per model checkpoint, and different sizes of the same family can carry different licenses.
- Llama's Community License permits commercial use up to 700 million monthly active users; above that, Meta requires a separate license it grants at its sole discretion.
- Qwen, Mistral, DeepSeek, GLM, OLMo, and Granite publish most current model sizes under Apache 2.0 or MIT (OSI-approved, no usage threshold), but several have also released specific checkpoints under more restrictive research, revenue-capped, or non-production licenses.
- Mistral's Modified MIT license, used on some checkpoints, blocks use once your organization's global consolidated monthly revenue exceeds $20 million — above that, you need a commercial license or Mistral's hosted service.
- Gemma uses Google's custom Gemma Terms of Use combined with a Prohibited Use Policy — a conduct-based restriction that OSI-approved open source licenses do not impose — and defines "Distribution" broadly enough to include hosting the model as an API or web service.
- Phi ships under a plain MIT license with no usage threshold; Falcon varies by version, from plain Apache 2.0 to a revenue-share clause above $1 million on newer releases.
- None of the ten licenses includes indemnification for the deploying business against third-party intellectual-property claims — all are distributed "as is," without warranty.
- "Open weight" is the more accurate term whenever a license carries a usage-scale limit, revenue cap, or conduct restriction — this covers Llama, Gemma, and any non-Apache/non-MIT checkpoint from the other families.
- This article is not legal advice. License texts change between model versions — verify the current license attached to your specific checkpoint and consult counsel before a commercial deployment.
Is This Legal Advice?
No — this guide is not legal advice, and it is not a substitute for reading the license text yourself. It summarizes, at a general level, how the commercial-use terms for ten widely deployed model families are currently structured, based on publicly published license documents. License terms change between model versions, vendors revise their terms without always renaming the license, and the specific facts of your deployment (revenue, user count, redistribution plans, jurisdiction) change how a given clause applies to you. Before any commercial deployment, read the exact license file shipped with the model checkpoint you intend to use and consult a qualified attorney familiar with software licensing and your jurisdiction.
Commercial LLM License Comparison
No two of these ten license structures are identical, and within several families different checkpoints carry different licenses. Use this table as an orientation, then verify against the current license file for your specific model before deployment.
Model | Commercial Use | License | Threshold | OSI Approved |
|---|---|---|---|---|
| Llama (Meta) | Yes, below 700M MAU | Llama Community License | 700M MAU/month | No |
| Qwen (Alibaba) | Yes (most sizes) | Apache 2.0 / custom (some) | None / some checkpoints | Yes (Apache sizes) |
| Mistral (Mistral AI) | Yes (most + capped some) | Apache 2.0 / Modified MIT | $20M/mo revenue (Mod. MIT) | Yes (Apache models) |
| Gemma (Google) | Yes, conduct limits apply | Gemma Terms of Use | None (conduct-based) | No |
| DeepSeek | Yes (current flagship) | MIT (current) / mixed (older) | None on MIT models | Yes (MIT models) |
| Phi (Microsoft) | Yes | MIT | None | Yes |
| GLM (Zhipu) | Yes (most models) | MIT / Apache 2.0 | None | Yes |
| OLMo (Ai2) | Yes | Apache 2.0 | None | Yes |
| Granite (IBM) | Yes | Apache 2.0 | None | Yes |
| Falcon (TII) | Yes (varies by version) | Apache 2.0 / TII Falcon License | $1M revenue share (newer) | Mixed by version |
"Commercial use allowed" does not mean every version of a model is commercially unrestricted — the license is set per checkpoint, not per brand. This table reflects the general license structure as commonly published; always confirm against the LICENSE file in the exact model repository you plan to deploy.
Can You Use a Local LLM Commercially?
Yes, under all ten license families covered here — but "commercial use" is not one legal question, it is several, and a license can answer them differently. Below is how the license categories above typically treat eight common business deployment scenarios; the exact answer still depends on the specific checkpoint's license text.
- Internal company chatbot (no external users) → Lowest-risk scenario under every family here. Usage thresholds and redistribution clauses rarely apply; the warranty disclaimer and absence of indemnification still apply regardless of internal-only use.
- Customer-facing chatbot or support tool → Check both the commercial-use permission and any conduct/prohibited-use terms (Gemma's Prohibited Use Policy, Llama's Acceptable Use Policy) — these govern what the model can be used to generate, not just who can use it.
- SaaS application built on the model → Check distribution and hosted-service provisions specifically. Gemma's Terms of Use define "Distribution" to include making the model available as a hosted service via API or web access — a SaaS product counts, even without shipping the weights themselves.
- API service reselling model access → Same hosted-service question as SaaS, plus confirm whether the license permits redistributing model outputs commercially at scale; Apache 2.0 and MIT checkpoints (most Qwen, Mistral, DeepSeek, GLM, OLMo, Granite sizes) impose no extra condition here.
- Mobile or desktop app embedding the model → Treat this as redistribution of the model itself, not just its outputs; attribution and pass-through conditions (Llama's "Built with Llama" credit, Gemma's Notice requirement) can apply even when the weights ship inside a compiled app.
- Fine-tuning a checkpoint on proprietary data → Generally permitted across all ten families for currently supported checkpoints, but derivative models can inherit the parent license's conditions — a fine-tuned Llama or Gemma derivative still carries the original attribution and conduct requirements forward.
- Redistributing model weights (including fine-tunes) → The scenario where license differences matter most. Apache 2.0/MIT checkpoints allow this with only a standard notice; Llama, Gemma, and non-Apache Qwen/Mistral/Falcon checkpoints attach pass-through attribution or conduct conditions to the redistributed copy.
- Selling a product or on-prem appliance built on the model → Combine the redistribution answer above with the usage-threshold question — a growing product could cross Llama's 700M MAU line or Mistral's Modified MIT $20M monthly revenue line well after initial deployment, so revisit the license at each major growth stage, not only at launch.
What Does "Open Weight" Mean vs. "Open Source"?
"Open weight" means the model's trained parameters are downloadable and runnable by anyone; "open source" in the OSI sense additionally requires the license to place no restriction on the field of use or the scale of the licensee's business. The Open Source Initiative's Open Source Definition specifically prohibits licenses that discriminate against fields of endeavor or against persons/groups — a usage-based threshold like Llama's 700 million MAU cap fails that test even though the weights themselves are freely downloadable.
Use "open weight" as the accurate description when a model's license includes any usage-scale threshold, a conduct-based prohibited-use policy, or a restriction tied to a specific business activity — this covers Llama, Gemma, and any Qwen or Mistral checkpoint released under a non-Apache license. Reserve "open source" for the specific model sizes genuinely released under an OSI-approved license text like Apache 2.0 or MIT, which several Qwen and Mistral checkpoints are.
This distinction matters commercially, not just semantically: a procurement or legal team evaluating "open source AI" as a category needs to know that the label does not guarantee unrestricted commercial use, and that the actual terms vary by the specific file being deployed, not by the vendor's marketing description.
📍 In One Sentence
A model is "open weight" if its parameters are freely downloadable; it is "open source" in the OSI sense only if the license also imposes no field-of-use or business-scale restriction, which excludes Llama, Gemma, and several Qwen/Mistral checkpoints.
💬 In Plain Terms
Downloadable does not mean unrestricted. Check whether the license has a usage cap or a rules-of-conduct policy attached — if it does, treat it as "open weight," not "open source," when your legal team asks the question.
LLM Licenses Explained: From Most Permissive to Most Restrictive
The ten families above map onto a small number of underlying license types — knowing which type you are dealing with tells you most of what you need before reading the fine print. From least to most restrictive on commercial use:
- MIT → Commercial use, modification, redistribution, and fine-tuning all permitted with only a copyright-notice requirement, no usage or revenue threshold. Phi and current DeepSeek and GLM checkpoints use this.
- Apache 2.0 → Functionally as permissive as MIT for a deploying business, with an added explicit patent grant. No usage threshold. Most current Qwen, Mistral, GLM, OLMo, Granite, and older Falcon checkpoints use this.
- Modified permissive licenses (e.g., Modified MIT with a revenue cap) → Behave like MIT below a stated threshold, then require a separate commercial license or the vendor's hosted service above it. Mistral applies this structure to some checkpoints, with a $20 million global monthly revenue cap.
- Custom business-scale licenses → Permit commercial use below a stated usage threshold (monthly active users, not revenue) and require a separately negotiated license above it. Llama's Community License (700 million MAU) is the clearest example.
- Custom terms with conduct restrictions → Permit commercial use but add a prohibited-use or acceptable-use policy governing what the model can be used to generate, independent of company size. Gemma's Terms of Use plus Prohibited Use Policy fits here; Llama layers this on top of its MAU threshold.
- Non-commercial / research-only licenses → Commercial deployment is not authorized without a separate agreement. Some historical Mistral, Qwen flagship, and older DeepSeek checkpoints have shipped under terms this restrictive — always confirm the specific checkpoint is not one of these before assuming Apache 2.0 or MIT applies.
- Each step down this list affects commercial use, fine-tuning, redistribution, SaaS deployment, and attribution differently — a checkpoint one step more restrictive than you assumed can turn a planned SaaS launch into a licensing conversation with the vendor.
What Are Llama's Commercial License Terms?
Meta distributes Llama under its own Llama Community License, which permits commercial use for most businesses but adds a usage-scale threshold that OSI-approved open source licenses do not have. If your product or service — including affiliates — had more than 700 million monthly active users in the month before the specific Llama version was released, the standard license does not authorize you to use it; you must request a separate commercial license directly from Meta, which Meta may grant or refuse at its sole discretion.
Below that threshold, the Community License permits commercial deployment, fine-tuning, and building derivative products. Recent versions have historically required attribution ("Built with Llama" or similar language) in specified circumstances, and have restricted using Llama or its outputs to train or improve a competing large language model outside the Llama family. Meta also maintains a separate Acceptable Use Policy prohibiting specific categories of use, similar in structure to Gemma's Prohibited Use Policy below.
The 700 million MAU threshold is a business-scale restriction, which is why the Llama Community License is not OSI-approved open source even though the weights are freely downloadable below that threshold — it fits the "open weight" category described above.
- Threshold: 700 million monthly active users of the licensee's (or its affiliates') products/services, measured in the month before the specific Llama version's release date
- Above the threshold: a separate license must be requested from Meta directly; Meta grants or denies at its sole discretion
- Attribution: recent Llama versions have required "Built with Llama" or similar credit in specified circumstances — check the version-specific license text
- Training restriction: some versions restrict using Llama outputs to train or improve a separate, competing large language model
- Acceptable Use Policy: a separate document lists prohibited use categories, enforced alongside the license itself
What Are Qwen's Commercial License Terms?
Alibaba releases most current Qwen model sizes under the Apache License 2.0 — a permissive, OSI-approved license with no usage-scale threshold, allowing commercial use, redistribution, and derivative models with only a standard notice-and-attribution requirement. This is the most business-friendly license among the four families for the sizes it covers.
The exception matters: Alibaba has also released specific Qwen checkpoints — historically some of the largest or newest flagship sizes — under a separate, more restrictive custom license (referred to in Qwen's own repositories as the Qwen License) that includes its own usage-scale threshold and, in some versions, a requirement to display the model's name in a derivative product. Because the license is attached to the checkpoint, not the "Qwen" brand as a whole, two different Qwen model sizes released the same week can carry two different licenses.
Before deploying any specific Qwen checkpoint commercially, open the LICENSE file in that exact model's repository (on Hugging Face or ModelScope) rather than assuming Apache 2.0 applies because other Qwen sizes use it.
- Most sizes: Apache License 2.0 — no usage threshold, OSI-approved, commercial use and redistribution freely permitted
- Some flagship/large checkpoints: a separate, custom Qwen License with its own commercial-use threshold — check per checkpoint
- Attribution: Apache 2.0 requires only the standard NOTICE/license retention; the custom Qwen License has, in some versions, required displaying the model name in derivative products
- Where to check: the LICENSE file shipped in the specific model's Hugging Face or ModelScope repository — never assume by brand name alone
What Are Mistral's Commercial License Terms?
Mistral AI ships most of its current open-weight releases under the Apache License 2.0 — the same permissive, OSI-approved terms Qwen uses for most sizes — but part of its catalog uses a Modified MIT License that adds a revenue cap Apache 2.0 does not have. As published in the LICENSE file attached to Mistral's Modified-MIT-licensed checkpoints, a licensee is not authorized to exercise the license's rights once their organization's global consolidated monthly revenue exceeds $20 million for the preceding month; above that line, the license directs you to either request a commercial license from Mistral AI or use the model through Mistral's own hosted service instead of self-hosting it. Verify this figure against the LICENSE file for your specific checkpoint — it is a per-model term, not a blanket company policy, and Mistral has attached it to some but not all of its current releases.
A model under Apache 2.0 can be deployed commercially, fine-tuned, and redistributed with only the standard notice requirement, regardless of company size. A model under the Modified MIT terms behaves the same way below the $20 million/month revenue line, then requires a separate commercial agreement above it. Some older Mistral checkpoints also remain under a research-only or non-production license with no commercial path at all — treat any Mistral checkpoint whose license file does not explicitly say Apache 2.0 or Modified MIT as non-commercial until confirmed otherwise.
Because Mistral has changed its licensing approach across releases, do not carry forward an assumption from a prior Mistral model you deployed — check the license attached to the specific checkpoint each time you evaluate a new one.
- Most current releases: Apache License 2.0 — commercial use, redistribution, and fine-tuning freely permitted, no revenue or usage threshold
- Some current checkpoints: Modified MIT License — same permissions below a $20 million/month global consolidated revenue threshold; above it, a commercial license or Mistral's hosted service is required
- Older/research checkpoints: some remain under a more restrictive research-only or non-production license requiring a separate commercial agreement regardless of revenue
- Default assumption: treat a Mistral checkpoint as non-commercial unless its LICENSE file explicitly states Apache 2.0 or Modified MIT
- Recheck each release: Mistral's licensing approach has changed over time — verify per model, not from memory of a prior release
What Are Gemma's Commercial License Terms?
Google distributes Gemma under its own Gemma Terms of Use, paired with a separate Gemma Prohibited Use Policy, rather than a standard OSI-approved open source license. Commercial use is generally permitted under these terms, but the license structure differs from Apache 2.0 or MIT in a way that matters to a procurement or legal review: it includes conduct-based restrictions (the Prohibited Use Policy bars specific categories of use, such as generating content that infringes rights or facilitates illegal activity) and Google reserves the right to restrict usage it believes violates those terms.
The Terms of Use define "Distribution" broadly: it covers any transmission or sharing of Gemma or a derivative to a third party, and explicitly includes making Gemma or its functionality available as a hosted service via API, web access, or any other electronic or remote means. In practice, that means offering a Gemma-powered SaaS product or API counts as Distribution under these terms even if you never hand a customer the model weights themselves, and any Distribution outside a hosted service must carry a Notice pointing back to the Gemma Terms of Use.
A field-of-use or conduct restriction of this kind is exactly the category of clause the OSI Open Source Definition prohibits, which is why Gemma — like Llama — belongs in the "open weight" category rather than "open source" in the strict sense, regardless of how the weights are marketed. Licensing terms can change between Gemma versions — read the terms page for the exact release you deploy, not a summary of a prior version, and do not assume a more permissive license applies to your checkpoint without confirming it in that release's own terms.
- License structure: custom Gemma Terms of Use plus a separate Gemma Prohibited Use Policy, not a standard OSI license
- Commercial use: generally permitted under current terms, subject to the Prohibited Use Policy's conduct-based restrictions
- "Distribution" includes hosted services: offering Gemma via API or web access as a SaaS product counts as Distribution under the Terms of Use, not just shipping the raw weights
- Enforcement: Google reserves the right to restrict usage it believes violates the Terms of Use or Prohibited Use Policy
- Version sensitivity: licensing terms can change between Gemma versions — read the terms page for the exact release you deploy, not a summary of a prior version
DeepSeek, Phi, GLM, OLMo, Granite, and Falcon Licenses
Beyond the four families with dedicated sections above, six other model families are common enough in local-LLM deployments to warrant their own license check. Each is summarized below; as with Llama, Qwen, Mistral, and Gemma, verify the LICENSE file for the exact checkpoint you deploy rather than relying on the family-level summary.
- DeepSeek — Current flagship DeepSeek releases ship under a plain MIT license: commercial use, modification, and redistribution permitted with no usage threshold. Some earlier DeepSeek checkpoints shipped under a separate, more restrictive non-standard license — treat any DeepSeek checkpoint that does not explicitly say MIT as unverified until you check its LICENSE file.
- Phi (Microsoft) — Phi models ship under a plain MIT license with no usage threshold or revenue cap, making it one of the least restrictive families on this page for commercial deployment, fine-tuning, and redistribution.
- GLM (Zhipu AI) — Most current GLM releases ship under MIT or Apache 2.0, both OSI-approved and threshold-free. As with Qwen and Mistral, confirm the license per checkpoint rather than assuming every GLM release uses the same terms.
- OLMo (Ai2) — OLMo releases ship under Apache 2.0 with no usage-count or revenue clause, alongside openly published training data and code — a comparatively transparent license structure among the families covered here.
- Granite (IBM) — IBM ships Granite models under Apache 2.0 across its current lineup, permitting commercial use, fine-tuning, and redistribution without a usage threshold.
- Falcon (TII) — Falcon's license depends heavily on the version. Some earlier Falcon releases ship under plain Apache 2.0. Newer releases use the TII Falcon License, an Apache-2.0-based license with an added acceptable-use policy and, in some published terms, a revenue-share clause once revenue attributable to Falcon or its derivatives exceeds roughly $1 million — check the specific license text attached to your Falcon checkpoint before assuming plain Apache 2.0 terms apply.
When Do Usage Thresholds Restrict Commercial Use?
A usage threshold converts a license from "commercial use permitted" to "separate license required" once your product crosses a specific scale — Llama publishes a hard, well-known number (700 million monthly active users), and Mistral's Modified MIT checkpoints and newer Falcon releases use a revenue-based version of the same idea instead of a user count. Qwen has used comparable threshold structures on some of its non-Apache checkpoints, historically framed around a combination of monthly active users and monthly revenue, though the exact figures are set per license and should be read directly rather than assumed.
Use case: a startup building a customer-facing product on a Llama-licensed model does not need to think about the 700 million MAU threshold at launch — it becomes relevant only if the company later reaches a scale most startups never hit, and legal counsel should revisit licensing at that growth stage, not before. The same logic applies to Mistral's $20 million/month revenue line and Falcon's roughly $1 million revenue-share trigger: they matter at a growth stage, not at initial deployment, but they belong on a compliance calendar rather than being ignored entirely.
For Apache 2.0- or MIT-licensed checkpoints (most Qwen, Mistral, DeepSeek, GLM, OLMo, and Granite sizes, plus Phi), there is no usage-scale or revenue threshold at all — the license permits commercial use regardless of company size or revenue. This is a meaningful differentiator when your business model anticipates high growth and you want to avoid revisiting the license question later.
What Attribution and Redistribution Rules Apply?
Attribution requirements range from "none beyond a standard license notice" (Apache 2.0, MIT) to "credit the model by name in specified circumstances" (Llama, some Qwen checkpoints). Redistribution — passing the model or a derivative of it to a third party — is permitted under all ten license structures for their current mainstream releases, but each attaches its own conditions.
Use Apache 2.0- or MIT-licensed checkpoints (most current Qwen, Mistral, DeepSeek, Phi, GLM, OLMo, and Granite sizes) when your deployment plan involves redistributing a fine-tuned derivative to customers or embedding the model in a product you resell — the attribution burden is limited to retaining the standard notice. Use extra care with Llama and Gemma derivatives: both require the derivative to carry forward the parent license's conditions (including, for Llama, the "Built with Llama" credit in specified circumstances, and for Gemma, the Notice requirement and adherence to the Prohibited Use Policy — which explicitly extends to hosting the model as an API or web service), and non-Apache/non-MIT Qwen, Mistral, and Falcon checkpoints can impose a similar pass-through requirement.
When NOT to assume redistribution is unrestricted: if your product embeds a fine-tuned Llama, Gemma, Modified-MIT Mistral, or newer Falcon derivative inside a SaaS product sold to enterprise customers, confirm whether the license's attribution, hosted-service, and pass-through conditions apply to your specific distribution model — API-only access, on-premises delivery, and bundled software each raise slightly different questions under these license texts.
Do These Licenses Include Indemnification or Warranty?
None of the ten license families offers indemnification to the deploying business, and all ten distribute the model "as is," without warranty of any kind — this is standard for open-weight and open-source licenses generally, not unique to any one vendor. If a third party later asserts an intellectual-property claim related to a model's training data or output, the license text does not obligate the model's publisher to defend or reimburse the business that deployed the model.
This differs from how some commercial, closed AI vendors structure enterprise agreements, where limited indemnification against IP claims is sometimes offered as a paid-tier feature. When evaluating open-weight models for a use case with elevated IP-litigation exposure, the absence of indemnification is a genuine business risk factor to weigh against the cost savings and control benefits of self-hosting — this is a business decision, not something a summary article can resolve for you.
Businesses in this position typically manage the risk through their own liability insurance, contractual risk allocation with their own customers, or by limiting deployment to internal/non-customer-facing use cases where third-party exposure is lower — none of which substitutes for a legal risk assessment specific to your situation.
Which License Fits Your Deployment Scenario?
Match your deployment scenario to the license structure that fits it, then verify the exact checkpoint's license file before committing engineering time to it.
- ✅ Startup or mid-size business, well under any usage or revenue threshold → Any of the ten families works from a threshold standpoint; prefer Apache 2.0- or MIT-licensed Qwen, Mistral, DeepSeek, Phi, GLM, OLMo, or Granite checkpoints if you want to avoid revisiting the license question as you grow.
- ✅ Building a resellable, white-labeled product with a fine-tuned derivative → Apache 2.0- or MIT-licensed checkpoints minimize attribution and pass-through obligations for the derivative you ship.
- ✅ Internal enterprise tool, no customer-facing redistribution → Attribution and redistribution clauses matter less; the usage threshold, revenue cap, and warranty disclaimer are still relevant regardless of internal-only use.
- ✅ SaaS or API product you plan to scale aggressively → Apache 2.0 or MIT checkpoints have no MAU or revenue ceiling to plan around; a Llama or Mistral-Modified-MIT choice means budgeting for a licensing conversation once you approach 700 million MAU or $20 million/month in revenue, respectively.
- ❌ Consumer product plausibly approaching hundreds of millions of monthly active users → Do not default to a Llama-licensed model without legal review of the 700 million MAU clause and a plan for what happens if you approach it; Apache 2.0/MIT alternatives avoid the question entirely.
- ❌ Any deployment with elevated IP-litigation exposure and no in-house legal risk tolerance → No license family here offers indemnification; evaluate whether a commercial vendor offering contractual indemnification is a better fit before defaulting to self-hosted open weights.
- If unsure: start with an Apache 2.0- or MIT-licensed checkpoint (available in most Qwen, Mistral, DeepSeek, Phi, GLM, OLMo, and Granite sizes) to eliminate the usage-threshold and revenue-cap questions, and route any commercial deployment plan through counsel before launch regardless of which family you choose.
What Are the Most Common Licensing Mistakes?
Most licensing problems come from treating a model family's brand name as a proxy for its license, rather than reading the specific file attached to the checkpoint being deployed.
- Mistake: Assuming "the whole family is Apache 2.0" (Qwen, Mistral, DeepSeek, GLM) as a blanket fact. Fix: Check the LICENSE file for the specific checkpoint — each of these vendors has released non-Apache/non-MIT models alongside their permissively licensed ones.
- Mistake: Ignoring the Llama 700 million MAU threshold or Mistral's $20 million/month revenue cap because the company is currently far below it. Fix: Note both thresholds in your compliance calendar and revisit them as part of any major growth-stage legal review, not only when you think you might already be close.
- Mistake: Treating "open weight" and "open source" as interchangeable when describing a model to your own legal or procurement team. Fix: Use "open weight" for Llama, Gemma, and any non-Apache/non-MIT checkpoint from the other families; reserve "open source" for genuinely OSI-licensed checkpoints.
- Mistake: Assuming a Gemma-powered SaaS product avoids Distribution obligations because you never hand customers the model weights. Fix: Read the Gemma Terms of Use's Distribution definition — it explicitly includes hosting Gemma as an API or web service.
- Mistake: Forgetting that redistribution and SaaS hosting are different legal questions from "can I use this model at all." Fix: Check the license's distribution, hosted-service, and pass-through clauses separately from its base commercial-use permission.
- Mistake: Ignoring fine-tuning-specific restrictions and assuming a derivative model inherits no obligations. Fix: Confirm whether the parent license's attribution, conduct, or pass-through conditions carry forward to your fine-tuned checkpoint — several do.
- Mistake: Assuming a license summary (including this one) is sufficient sign-off for a commercial launch. Fix: Route the final decision through counsel who reads the current license text for your exact use case — license text changes between versions and a stale summary can be wrong by the time you deploy.
- Mistake: Assuming any of these vendors will indemnify you if a third-party IP claim arises. Fix: Confirm explicitly that no such obligation exists in the license text, and plan your own risk mitigation (insurance, contractual allocation, or a commercial vendor with indemnification) accordingly.
- Mistake: Checking a model's license once at initial adoption and never again. Fix: Re-verify the license at every major version upgrade — vendors have changed licensing terms across releases before, on both the Apache/MIT side and the custom-license side.
What Is the Pre-Deployment License Checklist?
Work through this checklist before a commercial launch on any of the ten model families covered here.
- 1Identify the exact model checkpoint and its repository page
Why it matters: License terms are set per checkpoint, not per brand — you need the specific file, not a general summary. - 2Read the current LICENSE file (and any separate Terms of Use / Acceptable Use Policy) in full
Why it matters: Summaries — including this article — can miss version-specific clauses or be outdated by the time you deploy. - 3Check whether a usage-scale or revenue threshold applies and where your business currently sits against it
Why it matters: Llama's 700 million MAU clause, Mistral's $20 million/month revenue cap, and comparable Qwen and Falcon thresholds change the license outcome entirely above the line. - 4Confirm the attribution and hosted-service requirement for your specific distribution model (API-only, SaaS, on-prem, bundled)
Why it matters: Different distribution methods can trigger different attribution or pass-through obligations — Gemma's Terms of Use treat a hosted API as Distribution, for example. - 5Confirm there is no indemnification clause and document your own risk mitigation plan
Why it matters: None of the ten license families indemnifies the deploying business — this needs its own answer, not an assumption. - 6Route the final go/no-go decision through legal counsel familiar with your jurisdiction
Why it matters: This guide is not legal advice; the final sign-off should come from someone who can read the license against your specific facts. - 7Set a calendar reminder to re-check the license before your next major version upgrade
Why it matters: Vendors have changed licensing terms across model versions before — a license valid today is not guaranteed to remain unchanged at your next upgrade.
Frequently Asked Questions
Is this article legal advice about which LLM license to use?
No. This guide summarizes, at a general level, how the commercial-use terms for ten widely deployed model families are currently structured, based on publicly published license documents. It is not a substitute for reading the exact license text attached to your model checkpoint or for advice from a qualified attorney familiar with your jurisdiction and use case.
Can I use a local LLM commercially at all?
Yes — every model family covered in this guide (Llama, Qwen, Mistral, Gemma, DeepSeek, Phi, GLM, OLMo, Granite, Falcon) permits commercial use for at least some of its checkpoints. The question is not whether commercial use is allowed, but which conditions (usage thresholds, revenue caps, attribution, conduct restrictions) apply to the specific checkpoint and distribution method you plan to use.
Which LLM licenses allow commercial use with no threshold at all?
Apache 2.0 and MIT licenses have no usage-scale or revenue threshold. Phi ships under plain MIT; OLMo and Granite ship under plain Apache 2.0; most current Qwen, Mistral, DeepSeek, and GLM checkpoints also use one of these two licenses, though each of those four vendors has also released some checkpoints under more restrictive terms.
Can a small business use Llama commercially without contacting Meta?
Yes, if the business — including its affiliates — has fewer than 700 million monthly active users across its products or services. Below that threshold, the Llama Community License permits commercial use without a separate agreement. Above it, a separate license must be requested directly from Meta, which grants or denies it at its sole discretion.
Is Llama free for commercial use?
Yes, at no license cost, for businesses below the 700 million monthly active user threshold. Above that threshold, a separate license negotiated directly with Meta is required, and Meta sets its own terms for that agreement.
Are all Qwen models under the same license?
No. Alibaba releases most current Qwen model sizes under Apache License 2.0, which has no usage threshold, but has also released some checkpoints — historically some of the largest or newest flagship sizes — under a separate, more restrictive custom license. The license is attached to the specific checkpoint, not the Qwen brand as a whole, so check the LICENSE file in the exact model repository before assuming Apache 2.0 applies.
Is Qwen Apache 2.0 licensed?
Most current Qwen model sizes are, yes — but not all. Some flagship or research checkpoints have shipped under a separate, more restrictive Qwen License with its own usage terms. Check the specific checkpoint's LICENSE file rather than assuming by brand name.
Is Mistral always Apache 2.0 licensed?
No. Mistral AI ships most of its current open-weight releases under Apache License 2.0, but part of its catalog uses a Modified MIT License that blocks use once your organization's global consolidated monthly revenue exceeds $20 million — above that line you need a commercial license or Mistral's hosted service. Older research-only checkpoints exist too. Treat any Mistral checkpoint as non-commercial unless its LICENSE file explicitly states Apache 2.0 or Modified MIT.
Does Gemma's license count as open source?
Not in the strict OSI (Open Source Initiative) sense. Google distributes Gemma under a custom Gemma Terms of Use paired with a separate Prohibited Use Policy, which imposes conduct-based restrictions the OSI Open Source Definition does not permit in a genuinely open source license. Gemma is more accurately described as "open weight" — the parameters are freely downloadable, but the license carries conditions beyond a standard open source license.
Can I use Gemma commercially in a SaaS product?
Generally yes, subject to the Prohibited Use Policy's conduct-based restrictions. Note that the Gemma Terms of Use define "Distribution" to include making Gemma available as a hosted service via API or web access, so a Gemma-powered SaaS product falls under the same Terms of Use as shipping the weights directly.
What is the difference between "open weight" and "open source" for an LLM?
"Open weight" describes a model whose trained parameters are downloadable and runnable by anyone, regardless of the license terms attached. "Open source" in the OSI sense additionally requires the license to place no restriction on field of use or on the licensee's business scale. Llama, Gemma, and any non-Apache/non-MIT checkpoint from the other families are open weight but not open source in this strict sense; Apache 2.0- or MIT-licensed checkpoints meet both definitions.
Does Apache 2.0 allow commercial AI products?
Yes. Apache 2.0 is an OSI-approved permissive license with no usage-scale or revenue threshold — it permits commercial use, modification, redistribution, and building resellable products, subject only to retaining the standard copyright and license notice.
Can I fine-tune one of these models commercially?
Generally yes across all ten families for currently supported checkpoints, but the fine-tuned derivative can inherit the parent license's conditions. A Llama or Gemma fine-tune still carries forward attribution or conduct requirements from the base model; Apache 2.0/MIT fine-tunes carry only the standard notice requirement.
Can I redistribute an open-weight model or its weights?
It depends on the license. Apache 2.0- and MIT-licensed checkpoints permit redistribution with only a standard notice. Llama, Gemma, and non-Apache/non-MIT Qwen, Mistral, and Falcon checkpoints attach attribution or pass-through conditions — including, for Gemma, treating a hosted API as a form of Distribution — that the redistributed copy must also satisfy.
Do any of these vendors indemnify me if someone sues over training data or model output?
No. All ten license families distribute the model "as is," without warranty, and none includes an indemnification obligation for the deploying business against third-party intellectual-property claims. This is standard for open-weight and open-source licenses generally. Businesses with elevated IP-litigation exposure should weigh this against commercial vendors that sometimes offer indemnification as a paid enterprise feature.
Do I need to credit the model publisher when I use their model?
It depends on the specific license. Apache 2.0- and MIT-licensed checkpoints require only a standard notice/license retention, not visible product credit. Llama has, in specified circumstances across recent versions, required a "Built with Llama" credit; Gemma requires a Notice for non-hosted Distribution, and some non-Apache Qwen checkpoints have carried their own attribution or naming conditions. Check the current license text for the specific checkpoint you deploy.
Does self-hosting one of these models change the license terms?
No — the license governs the model itself, not where you run it. Self-hosting versus using a cloud-hosted API of the same model does not change the underlying license conditions (usage thresholds, revenue caps, attribution, or the absence of indemnification), though it does change other factors like data residency and your own infrastructure obligations, which are separate legal considerations from the model license itself.
How often do these license terms change?
They change more often than most businesses expect — every vendor covered in this guide has revised license terms across model versions, and a company's licensing approach for its newest release does not automatically apply retroactively to older checkpoints still in use. This is why this article is set to a semi-annual refresh cycle and why the checklist above recommends re-verifying the license before every major version upgrade rather than relying on a one-time review.
Where Can You Find the Primary License Texts?
- Llama Community License (llama.com/llama3_1/license) — the official license text, including the 700 million MAU clause; always check the version-specific page for the model you deploy
- Qwen model repositories (huggingface.co/Qwen) — each checkpoint's repository includes its own LICENSE file; do not assume Apache 2.0 applies without checking
- Mistral AI model releases (mistral.ai/news) and Hugging Face repositories — check the LICENSE file attached to the specific model card, including the Modified MIT revenue-cap terms where applicable
- Gemma Terms of Use (ai.google.dev/gemma/terms) and Gemma Prohibited Use Policy (ai.google.dev/gemma/prohibited_use_policy) — the two documents that together govern Gemma use
- DeepSeek model repositories (huggingface.co/deepseek-ai) — check the LICENSE file per checkpoint; current flagship releases are MIT
- Phi model repositories (huggingface.co/microsoft) — Microsoft's Phi models ship under a plain MIT license
- GLM model repositories (huggingface.co/zai-org) — check the LICENSE file per checkpoint; most current releases are MIT or Apache 2.0
- OLMo model repositories (huggingface.co/allenai) — Ai2's OLMo releases ship under Apache 2.0
- Granite model repositories (github.com/ibm-granite) — IBM's Granite models ship under Apache 2.0
- Falcon LLM license page (falconllm.tii.ae) — check the version-specific license; terms have changed between Falcon releases
- Open Source Initiative — Open Source Definition (opensource.org/osd) — the reference standard used in this guide to distinguish "open weight" from "open source"