What counts as an open model, and what the licenses allow

People use "open" for two different things when they talk about AI models. The Open Source Initiative (OSI) has a formal definition of open source AI, and its bar is high. The everyday meaning, and the one tryingopen uses, is open weights: the trained model is published, anyone can download it, and a license says what you may do with it.

OSI's Open Source AI Definition

OSI publishes the Open Source AI Definition. As of October 2026 it is still version 1.0, released on October 28, 2024. It asks for four freedoms:

• Use the system "for any purpose and without having to ask for permission."
• Study how it works and inspect its components.
• Modify the system for any purpose, "including to change its output."
• Share it with others, with or without changes.

To use those freedoms you need what the definition calls the preferred form for making modifications. It has three parts:

• Data information: "Sufficiently detailed information about the data used to train the system," enough for a skilled person to build a substantially equivalent system.
• Code: "The complete source code used to train and run the system."
• Parameters: the weights and other configuration settings.

The definition says that "'Open Source models' and 'Open Source weights' must include the data information and code used to derive those parameters." OSI's page on open weights puts it this way: "Open Weights alone fall short of this because they do not provide the underlying training process, code, or comprehensive data details required for full-fledged use, study, modification, and sharing."

The models OSI checked

While writing the definition, OSI tested a set of models against it. Its FAQ lists the ones that passed: Pythia (EleutherAI), OLMo (AI2), Amber and CrystalCoder (LLM360), and T5 (Google). Llama2, Grok, Phi-2 and Mixtral failed, because they "lack required components and/or their legal agreements are incompatible with the Open Source principles." OSI adds that these results "are not certifications of any kind."

Our picker uses the everyday meaning from the first paragraph, open weights, with two exceptions covered next.

What the picker means by open

Open the model picker and you'll see this note:

23 of these 24 are open: anyone can download and run them. The other is announced as open, with the weights still to come.

The one left out of the count is Mistral Large 4, which carries a "Weights not out yet" badge. Mistral announced it on October 6, 2026 and wrote: "We will release the weights by the end of the month." As of October 14, 2026, the weights have yet to appear on Hugging Face, and the announcement names no license.

Qwen3.8 Flash, counted among the 23, is Alibaba's hosted production model. The downloadable model behind it is Qwen3.8-Flash-Next, whose card says "Qwen3.8-Flash is the official version based on Qwen3.8-Flash-Next with more production features." Flash-Next has 125B parameters with 6B active, and Alibaba publishes it under the qwen-community-1.0 license.

Because the weights are published, we run two of the picker's models ourselves. We serve 4-bit builds of LFM2.5-VL 3B and Gemma 4 E2B on our own Jetson Orin Nano boards (how that works).

The licenses in the catalogue

Here is how the 24 models in the picker break down by license as of October 2026, going by the Hugging Face cards:

Apache 2.0 and MIT let you use, change and redistribute the weights, commercially too, as long as you keep the copyright and license notice. Apache 2.0 also grants a patent license and requires you to mark files you changed. Both apply the same terms at any company size.

OpenMDW-1.1, the license NVIDIA uses for both Nemotron models, grants permission "free of charge, to deal in the Model Materials without restriction." Anyone who redistributes the model must keep the license and the original notices, and your rights end if you sue claiming the model infringes a patent or copyright (unless you are answering a suit brought against you first). It sets no conditions on outputs.

Custom licenses with business conditions

Four custom licenses in the catalogue add conditions for some businesses. Three of them apply only past a revenue or user-count line.

lfm1.0 (Liquid AI) covers LFM2.5-VL 3B, our default model, and LFM2.5 24B. It also covers d1-3B, the router that picks between our small model and the cloud (how d1 routes). The license allows commercial use while your company's annual revenue stays below $10 million. A company at or above that line needs a separate commercial license from Liquid AI. The license page says "Research and Nonprofit Usage is specifically allowed with no thresholds."

glm-5.3 (Z.ai) covers GLM 5.3. It allows commercial use, modification and redistribution. A company that runs a model-as-a-service business, with group revenue above $10 billion in any 12 months, must "pass Z.AI's security review" before using the model commercially. Its smaller sibling, GLM 5.3 Flash, is MIT.

qwen3.8-max (Alibaba) covers Qwen3.8 2.4T. A model-as-a-service or AI work assistant business with group revenue above $50 million in any 12 months needs a separate license from Qwen before commercial use. Internal use that keeps the model and its outputs away from third parties is exempt. Any product built on it with more than 100 million monthly active users or $20 million in monthly revenue must display the model name prominently in its interface.

qwen-community-1.0 (Alibaba) covers Qwen3.8-Flash-Next and sets the same rules, except that a model-as-a-service or AI work assistant business of any size needs Qwen's separate license before commercial use.

A company under $10 million in annual revenue, with no product above 100 million monthly active users, clears every revenue and user-count line above.

Checking a license yourself

The info button on each picker row opens a details panel, which shows the maker's announcement for 11 of the 24 models. The license lives on the model's Hugging Face card, as a tag near the top and as a LICENSE file in the repository. For Apache 2.0 and MIT the tag tells you enough. For a custom license, read the file: GLM 5.3's security review requirement is spelled out in its LICENSE file. Start with the cards for GLM-5.3 and Qwen3.8-2.4T-A95B.

To run a smaller model on your own computer and chat with it from tryingopen, see using Ollama, LM Studio, llama.cpp or Jan from Self Chat. To choose one by what you need, see the tour of the 24 models in the picker.