Open-Weight Model Licenses & Commercial Use: October 2026 Enterprise Guide

In short: An open-weight model lets you download the numerical parameters (weights) and run them on your own servers, but commercial usage rights come from the license text rather than the download link. Liquid AI’s LFM3, Meta’s Muse Spark Open, and Mistral Large 4 all shipped in October 2026 as open-weight models, yet their revenue caps, Model-as-a-Service (MaaS) clauses, and acceptable-use policies differ sharply. This article is not legal advice; the license summaries reflect the texts as published on October 30, 2026. Always review the license with your legal counsel for your specific use case.

This deep dive accompanies our Week 44 (Oct 22–29, 2026) Enterprise AI Briefing. Disclosure: RuneLab provides on-premise model appliances (RuneBox) and corporate training on open-model selection and AI governance; RuneLab does not provide legal advice.

In an open-weight release, the numerical parameters learned during training are published so engineering teams can download, self-host, quantize, and fine-tune the model inside their own infrastructure. However, downloadable weights do not imply unconditional commercial rights.

“Open source” has a stricter definition in software licensing that prohibits discrimination against fields of endeavor or specific groups of users. Custom model licenses that impose annual revenue caps, industry exclusions, or Model-as-a-Service restrictions generally fall outside that definition. Two models announced as “open-weight” in the same week can carry completely different legal obligations for an enterprise or holding group.

Liquid AI announced the LFM3 family on October 27, 2026, starting with six 2B-class multimodal models. The weights ship under the LFM Open License v1.0, and the license text sets these terms:

• Built on the Apache 2.0 structure, granting rights to use, modify, and distribute the model and derivatives.
• Commercial revenue threshold ($10M USD): The license defines the threshold as $10 million USD or more in annual revenue. Entities at or above $10M USD in annual revenue (where the license’s “Legal Entity” definition includes entities under common control) are excluded from the free commercial grant and must obtain a paid commercial license from Liquid AI.
• Group and subsidiary calculation: Revenue is evaluated across the “Legal Entity,” which includes controlling, controlled, and common-control affiliates. A subsidiary whose standalone revenue is below $10M may still fall above the threshold if its parent group has $10M+ in annual revenue; confirm with counsel.
• Non-profit exemption: The threshold applies only to commercial use; non-commercial or research use by qualifying non-profit organizations is exempt.
• No copyleft requirement: You do not have to open-source fine-tuned derivatives, provided you retain copyright, attribution, and NOTICE files when distributing.
• Automatic termination and deletion: Upon breach, the license terminates automatically and requires ceasing use and deleting all copies of the model and derivatives.

For hardware sizing and edge deployment patterns, see our Liquid AI LFM3 Enterprise Guide.

Meta released Muse Spark Open on October 15, 2026. The Meta AI Research announcement describes an 816B total / 42B active parameter multimodal model with a 1M-token context window. The weights ship under the Muse Model License, published on the model’s Hugging Face page:

• Permits use, copying, modification (including fine-tuning), distillation, derivative creation, distribution, and sale.
• Required attribution badge: Every product or service using the software or its derivatives must prominently display Built with Muse Spark Open in the user interface, documentation, website, or settings menu.
• Broad derivative definition: Any model trained, modified, or distilled using Muse Spark Open or its outputs is treated as a derivative under the license.
• $20M gross-revenue MaaS threshold: If the licensee or its affiliates provide third parties with API or hosted access to the model and the group’s total gross revenue exceeds $20 million USD in any 12-month period, the licensee must notify Meta and execute a separate written agreement before offering the model as part of that service. Internal use and embedded product features are excluded from this MaaS trigger.
• Exhibit A Usage Policy: States the model is not designed for users under 18 and prohibits unlicensed practice of regulated professions (finance, law, healthcare), military/ITAR applications, critical infrastructure operation, and misrepresenting machine outputs as human-generated.

Mistral AI released the open weights of Mistral Large 4 on October 29, 2026 under the Mistral Ecosystem Open-Weight License. The LICENSE file in the Hugging Face repository sets these terms:

• Permits using, copying, modifying, distributing, selling, running, post-training, and creating derivatives of the model.
• Two core conditions: Keep the copyright and permission notice (2026 - Mistral) in all copies; and if the licensee or any of its affiliates offers the model to third parties as a Model-as-a-Service (MaaS), any non-internal commercial use requires a separate commercial license from Mistral AI. There is no revenue threshold.
• How MaaS is defined: Giving a third party access to the model or its derivatives for inference or post-training (for example via an API or hosted endpoint) with meaningful control over inputs, parameters, or training data.
• Internal use and embedded features: Internal use (where the model is not exposed to third parties) is permitted, and end-user products that embed the model solely to power specific product features do not count as MaaS. Whether serving sister companies inside a holding group counts as “third-party” access is not explicitly defined in the license; verify with counsel.

For GPU cluster sizing and TCO analysis, see our Mistral Large 4 On-Premise Guide.

For comparison, Google DeepMind’s EmbeddingGemma 2 lists Apache 2.0 on its Hugging Face card (meaning no revenue cap or MaaS restriction), yet the same model card notes that usage must comply with the Gemma Prohibited Use Policy. Always review both the repository LICENSE file and the model card before approving an open-weight artifact.

Model	Release Date	License Name	Revenue / MaaS Condition	Attribution	Usage Policy Note	Primary Source
LFM3 (Liquid AI)	Oct 27, 2026	LFM Open License v1.0	Separate license required for commercial use at or above $10M USD annual revenue (includes common-control group under “Legal Entity”); no separate MaaS clause	Keep license copy, modification notices, copyright, and NOTICE	No separate acceptable-use exhibit in license text	LFM License
Muse Spark Open (Meta)	Oct 15, 2026	Muse Model License	MaaS providers whose group gross revenue exceeds $20M USD over 12 months must notify Meta and sign a separate agreement; internal use and embedded features excluded	Display “Built with Muse Spark Open” in UI or docs	Exhibit A Usage Policy restricts under-18 use and unlicensed regulated professions	HF Model Card
Mistral Large 4 (Mistral AI)	Oct 29, 2026	Mistral Ecosystem Open-Weight License	If licensee or affiliate offers the model as third-party MaaS, non-internal commercial use requires a separate license; no revenue cap; internal use and embedded product features permitted	Retain copyright and permission notice in all copies	Model card prohibits infringing third-party IP or rights	LICENSE
EmbeddingGemma 2 (Google)	2026	Apache 2.0	None	Standard Apache 2.0 notices	Model card references Gemma Prohibited Use Policy	Model Card

1. Does a revenue cap apply to our corporate group? Does the license aggregate parent and sister-company revenue under common control?
2. Does our planned architecture count as MaaS? Are we exposing raw endpoints to external clients or affiliates, or embedding the model behind a specific application workflow?
3. Where must attribution notices appear? Who owns maintaining NOTICE files and UI badges across releases?
4. Do acceptable-use exhibits restrict our sector? Do healthcare, finance, legal, defense, or under-18 clauses apply to our deployment?
5. How are fine-tuned or distilled derivatives governed? If we generate synthetic training data with the model, does the output taint downstream student models?
6. What are the governing law and termination obligations? Does a breach trigger automatic termination and mandatory deletion of fine-tuned weights?

Even when a model license permits commercial use, procurement rules and data protection laws apply independently. For example, on October 20, 2026, the U.S. Department of Homeland Security published a Federal Register notice excluding DeepSeek products and third-party services built with DeepSeek implementations from U.S. federal civilian agency IT systems. Similarly, running an open-weight model on-premise does not replace KVKK compliance: prompt logging, retention windows, and personal data inside fine-tuning datasets must be reviewed with your data protection counsel.

• Check closely if your group’s annual revenue is at or above $10M USD: LFM3 requires a commercial license for commercial use.
• Check closely if you expose API or hosted model endpoints to customers or partners: Muse Spark Open and Mistral Large 4 MaaS definitions may apply (with no revenue threshold on Mistral).
• Check closely if you train smaller models on synthetic outputs: Muse Spark Open’s derivative definition covers models trained or distilled from its outputs.

Teams standardizing open-weight governance can explore RuneLab’s Open-Weight Model Selection & Evaluation and AI Governance courses, RuneBox on-premise appliances, and our Apache 2.0 Turkish models on RuneLab Models.

Once an engineering team downloads an open-weight checkpoint onto an internal GPU node and passes initial Turkish accuracy benchmarks, it is easy to assume deployment is purely a DevOps task. In practice, pre-production legal and procurement reviews consistently surface four concrete governance scenarios.

Enterprises in banking, energy, retail, and manufacturing frequently house their AI engineering teams inside a dedicated technology or R&D subsidiary whose standalone revenue is modest.

Because licenses like LFM Open License v1.0 define “Legal Entity” to include parent and sister companies under common control, the parent group’s revenue can push the subsidiary above the $10M USD threshold.

Engineering leads should verify group-level revenue eligibility with legal counsel before investing weeks in fine-tuning a revenue-capped base model.

Running an open-weight model behind an internal employee workflow carries a very different legal profile from exposing that same model to external customers, dealers, or suppliers via an API endpoint.

Under the Mistral Ecosystem Open-Weight License, internal use is permitted, whereas offering the model as a hosted service where third parties control inputs or parameters triggers a separate commercial license requirement regardless of revenue.

System architects should document early whether a model is strictly embedded inside a specific application feature or exposed as a general-purpose inference endpoint.

To reduce inference latency and hardware cost, teams often prompt a large open-weight teacher model to generate synthetic training pairs and distill those capabilities into a compact 2B–9B student checkpoint.

Under licenses such as Meta’s Muse Model License, models trained or distilled using the software’s outputs are explicitly classified as derivatives, inheriting attribution badges and acceptable-use restrictions.

Always check the teacher model’s derivative clause before generating synthetic datasets for downstream commercial models.

When multiple product squads pull checkpoints directly from public repositories, tracking attribution requirements and acceptable-use exhibits across production services quickly becomes error-prone.

Establishing a curated internal model registry that records the exact license file, model-card policy, and required NOTICE text for every approved checkpoint prevents last-minute launch delays.

No. An open-weight model lets you download the trained parameters to run on your own hardware, while commercial usage, redistribution, and derivative rights depend on the specific license text. Licenses with revenue thresholds or Model-as-a-Service (MaaS) restrictions generally do not meet the strict Open Source Definition.

It depends on the license. As of October 30, 2026, LFM Open License v1.0 requires a paid commercial license for entities at or above $10M USD in annual revenue (including common-control group entities under its “Legal Entity” definition). Muse Spark Open requires a separate agreement for MaaS providers exceeding $20M USD in group gross revenue. Mistral Large 4 permits internal use, but requires a separate license for non-internal commercial use if you or an affiliate offer the model as third-party MaaS. Always confirm with legal counsel.

Yes. LFM Open License v1.0, Muse Model License, and the Mistral Ecosystem Open-Weight License contain no copyleft obligation requiring you to publish fine-tuned weights publicly. However, the underlying attribution, MaaS, revenue-cap, and acceptable-use conditions continue to apply to your fine-tuned derivatives.

LFM Open License v1.0 requires a separate commercial license from Liquid AI for commercial use by entities with $10 million USD or more in annual revenue. The threshold applies only to commercial use (non-commercial or research use by qualifying non-profits is exempt). Because “Legal Entity” includes entities under common control, a subsidiary may fall above the threshold if its corporate group has $10M+ in annual revenue; confirm with counsel.

Mistral Large 4’s license permits internal use. If an organization or its affiliates offers the model to third parties as a Model-as-a-Service (MaaS), giving third parties API or hosted access with meaningful control over inputs, parameters, or training data, any non-internal commercial use requires a separate commercial license from Mistral AI. End-user products that embed the model solely to power specific features do not count as MaaS, and there is no revenue threshold. Whether serving sister companies inside a holding group counts as “third-party” access is not defined in the license; ask your legal counsel.