DeepSeek permits developers to modify and sell services built with its R1 model. Meta adds branding requirements to its Llama 4 license and requires separate permission for some very large businesses. OpenAI distributes its gpt-oss models under Apache 2.0, a widely used software license.
All three make model weights available for download. Those are the numerical settings produced during training, which developers need to run a model themselves. Downloading weights allows an organization to operate the software on computers it controls, subject to the accompanying license.
What developers can change and redistribute
When DeepSeek released R1 on January 20, 2025, it made the code and weights available under the MIT license. Its license text allows use, copying, modification, distribution and sale, provided the required copyright and permission notice remains with copies. DeepSeek also said developers could use R1’s answers to train or improve other models.
The company separately released six smaller models trained using answers from R1. Its technical repository identifies the underlying Qwen and Llama models and their original licenses. Several start from Qwen models licensed under Apache 2.0; two use Llama models, which have Meta’s own terms.
Meta’s Llama 4 license, dated April 5, 2025, allows modification and commercial distribution but requires recipients to include its agreement and display “Built with Llama” with products using the materials. A distributed model trained or improved with Llama materials or outputs must have a name beginning with Llama.
Another clause covers companies whose products or services, together with those of their affiliates, exceeded 700 million monthly active users in the month before Llama 4’s release. They must ask Meta for a separate license. Meta reserves the decision whether to grant one.
Computers to run them
OpenAI released gpt-oss-120b and gpt-oss-20b on August 5, 2025. Both use Apache 2.0. OpenAI said developers could run the larger model on one graphics processor with 80 gigabytes of memory. The smaller version fits in 16 gigabytes, making it usable on some consumer computers. The company described specialized training and operation on customers’ own infrastructure among its intended uses.
Running a larger downloadable model can require considerably more equipment. In January 2025, Nvidia packaged DeepSeek R1 for its NIM software service, which companies can install on their own systems. Nvidia’s example for the full R1 model used a server containing eight H200 graphics processors.
