Open licenses: free to build on, but hosting needs a deal

Six of the biggest 2026 open-weights licenses let you build products freely but require a separate agreement or review from businesses that sell hosted access. Here is how the thresholds differ.

  • Analysis
  • 3 min read
  • Published
  • Updated

Read the licenses attached to the largest open-weights releases of 2026 and a pattern appears. You may use the model, fine-tune it and sell products built on it. What several developers now restrict is a specific business: the one that sells other people access to the model. This article summarizes what the licenses in our catalog say. It is a reading aid, not legal advice.

The pattern in one table

License Model in our catalog Who needs a separate deal or review Threshold
Kimi K3 License Kimi K3.1 Pro Businesses that sell API or fine-tuning access Combined revenue above US$20 million in any 12 months
Muse Model License Muse Spark Open Businesses that offer the model to third parties through an API Group revenue above US$20 million in any 12 consecutive months
Qwen3.8-Max License Qwen3.8-Max Model-as-a-service or “AI work assistant” businesses Combined revenue above US$50 million in any 12 months
Qwen Community License 1.0 Qwen3.8-Flash-Next Any model-as-a-service or AI work assistant business None: any size
Mistral Open-Weight License Mistral Large 4 Any business that gives third parties hosted access None: any size
GLM-5.3 License GLM-5.3 Model-as-a-service businesses must pass a Z.AI security review Combined revenue above US$10 billion over any 12 months

Two things stand out. The thresholds range from nothing at all (Mistral and Qwen Community) to US$10 billion (GLM-5.3), so the same kind of business is treated very differently depending on the model. And the strictest rule is not always on the flagship: Qwen3.8-Flash-Next, a much smaller model, has a stricter hosting rule than the far larger Qwen3.8-Max.

The newest case is stricter still. MiniMax H3-Image, released on 10 November, needs MiniMax’s authorization for any hosted service offered to third parties, whatever its revenue, and for any business above US$1 million a year. Its license also excludes the United States, the EU, the UK and South Korea (a territory restriction), so it sits outside the pattern in the table.

The idea has reached speech models too. The Fish Audio Community License behind Drama 3 never covers a hosted text-to-speech or voice-cloning service, and it ends free commercial use at US$500,000 of yearly revenue.

What counts as hosting

The licenses use different words, but they describe the same activity: giving third parties access to the model’s inference or fine-tuning, for example through an API or a hosted endpoint. Kimi K3 talks about selling API or fine-tuning access, Mistral about giving third parties access for inference or post-training, and the GLM license about access with real control over inputs, parameters or training data.

Several things are explicitly not hosting:

  • using the model internally, at any scale;
  • embedding it as a feature of your own product, such as a chat helper or a document tool;
  • relaying requests to a model that someone else hosts.

Qwen goes one step further than the others. Its licenses also cover “AI work assistant” businesses, meaning an independent coding or office-productivity product, so a company building a standalone assistant on top of Qwen needs to read those clauses even if it does not sell API access.

What it means for you

  • Individuals, hobbyists and researchers: none of this affects you.
  • Companies using a model inside their own products or tools: mostly unaffected. Check the credit lines instead. Muse Spark Open requires everyone to display “Built with Muse Spark Open”, while Kimi K3 and Qwen3.8-Max ask for the model name only on very large products.
  • API providers and hosting platforms: you are the audience for these clauses. Check the threshold, or whether there is one, before you offer the model.
  • Standalone AI assistants: if you build on Qwen, check the “AI work assistant” language.

The permissive alternatives

Not every developer takes this route. Inkling, Reflection Beam and Command A+ are Apache 2.0, DeepSeek V4.1 Flash, Ling-3.1-flash and GLM-5.3-Flash are MIT, and Nemotron 3.5 uses OpenMDW-1.1. None of these has a hosting clause. The GLM family shows the split inside one developer: GLM-5.3 uses the custom license above, while its smaller sibling GLM-5.3-Flash is MIT.

The models you can use commercially page sorts the catalog by license type, and AI licenses explained has a checklist to run before you ship.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Kimi K3.1 Pro Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster
Muse Spark Open Meta 816B 42B active 1M Muse Model License (Conditional) 563 GB GPU server
Qwen3.8-Max Alibaba Qwen 2.45T 95B active 1M Qwen3.8-Max License (Conditional) 1.69 TB Multi-node cluster
Qwen3.8-Flash-Next Alibaba Qwen 180B 6B active 256K Qwen Community 1.0 (Conditional) 124 GB Big-memory desktop
Mistral Large 4 Mistral AI 1.05T 52B active 1M Mistral Open-Weight (Conditional) 671 GB GPU server
GLM-5.3 Z.ai (Zhipu AI) 753B 40B active 1M GLM-5.3 License (Conditional) 520 GB GPU server
MiniMax H3-Image MiniMax 33.1B – MiniMax H3-Image Community (Conditional) 155 GB Big-memory desktop
Drama 3 Fish Audio 35.9B 3B active – Fish Audio Community 1.0 (Conditional) 25 GB Gaming GPU