Key takeaways
- Open weights means you can download and run a trained model; the Open Source Initiative's definition also requires training code and a detailed description of the training data.
- DeepSeek-V4-Pro (MIT), gpt-oss-120b and small Qwen3.8 models (Apache 2.0) are open weights under OSI-approved licenses, while Llama 4 and Qwen3.8's largest model use custom licenses with user and revenue thresholds.
- Alibaba's Qwen4-Max, released November 16, 2026, is API-only, and Alibaba says smaller Qwen4 models are coming to open source without giving a date.
Open weights means you can download a trained model's parameters and run them yourself. Open source AI, as the Open Source Initiative defines it, means you also get the training code and enough information about the training data to rebuild the model, under terms that let anyone use, study, change and share it for any purpose. Most models called "open source" today, from DeepSeek, Qwen, Meta and OpenAI, are open weights. Only a few, such as Ai2's Olmo, meet the full definition.
The difference decides what you can do. Open weights let you run a model on your own hardware, keep data private and fine-tune it. Only full openness lets you see what the model learned from or train your own version from scratch.
What open weights means
A model's weights, or parameters, are the billions of numbers it learned during training. With them and some inference code, you can run the model on your own servers, shrink it with quantization so it fits on smaller GPUs, and fine-tune it on your own data.
The OSI's page on open weights calls them "the final weights and biases of a trained neural network" and notes what they leave out: the training code and the training dataset. Without those, you can use the model but not reproduce it.
Open weights also says nothing about the license. Some models come under standard open source licenses with no strings. Others come under a license the company wrote, with user caps, naming rules and usage policies.
What the Open Source AI Definition requires
The OSI, which maintains the definition of open source software and approves open source licenses, released version 1.0 of the Open Source AI Definition on October 28, 2024. It carries the four freedoms of open source software over to AI: you must be able to use the system for any purpose without asking permission, study how it works, modify it for any purpose, and share it with or without changes.
To make those freedoms real, it asks for three things:
- Data information. A description of the training data detailed enough that a skilled person could build a substantially equivalent system: where the data came from, how it was selected and filtered, and where to get any public or third-party data.
- Code. The full code used to train and run the model, under an OSI-approved license.
- Parameters. The weights, under OSI-approved terms.
The definition does not require publishing the training data itself. The OSI FAQ says it wants open source AI to exist "in fields where data cannot be legally shared, such as medical AI." Open data must be shared, public or purchasable data must be pointed to, and data that cannot be shared must be described in detail.
When the OSI tested models against the definition, the FAQ says Pythia, OLMo, Amber, CrystalCoder and T5 passed, and Meta's Llama 2 did not. The OSI adds that these results are "not certifications of any kind."
Open weights vs open source, side by side
| Open weights | Open source AI (OSI) | |
|---|---|---|
| Download, run and fine-tune | Yes | Yes |
| License | Any: standard or custom | OSI-approved terms only |
| Commercial use | Depends on the license | Yes, for any purpose |
| Training code | Usually not released | Required |
| Training data | Usually not described | Detailed description required; open data must be shared |
| Retrain from scratch | No | In principle, yes |
| Examples | DeepSeek V4, Qwen3.8, Llama 4, gpt-oss | Olmo, Pythia |
What real model licenses let you do
The license is where open-weight models differ most. This is what the best-known ones say, as of November 16, 2026.
Apache 2.0 and MIT: open weights with no strings
DeepSeek-V4-Pro, with 1.6 trillion parameters of which 49 billion are active per token, ships its repository and weights under the MIT license. DeepSeek's R1 model card says the license allows commercial use, any changes, and distillation, meaning training other models on its outputs. No training data or training code is linked. More in Is DeepSeek open source?
OpenAI's gpt-oss-120b, with 117 billion parameters and 5.1 billion active, uses Apache 2.0, as do Alibaba's smaller Qwen models such as Qwen3.8-27B. Apache 2.0 and MIT are OSI-approved licenses, so these weights meet the "parameters" part of the definition. They fall short on code and data information, which is why they count as open weights.
Custom licenses: open weights with conditions
Meta's Llama 4 Community License, effective April 5, 2025, adds conditions:
- A company whose products had more than 700 million monthly active users in the previous month must ask Meta for a license, which Meta may grant "in its sole discretion."
- Anyone who distributes Llama or a product built on it must show "Built with Llama," and a distributed model trained or fine-tuned with Llama must start its name with "Llama."
- All use must follow Meta's Acceptable Use Policy.
Limits on who may use a model, and for what, conflict with the "any purpose" freedom.
Alibaba uses both kinds. Its largest open model, Qwen3.8-2.4T-A95B, ships under its own Qwen3.8-Max License. A product with over 100 million monthly active users or over $20 million in monthly revenue must display the model's name. A company selling model access or an AI coding or office assistant with over $50 million in revenue in 12 months needs a separate license. Internal use is exempt.
Fully open: what Ai2's Olmo releases
Olmo, from the Allen Institute for AI (Ai2), is the clearest example of a model that meets the definition. Ai2 calls it a "fully open language model" and releases:
- Weights for the Olmo 3 family at 7 billion and 32 billion parameters, under Apache 2.0.
- Training data: the Dolma 3 pretraining mix and the post-training datasets.
- Code: OLMo-Core for pretraining, Open-Instruct for post-training and OLMo-Eval for evaluation.
- Checkpoints from each training stage.
The Olmo 3.1 32B Think model card says training logs are still "coming soon." Fully open models are also smaller: Olmo 3 tops out at 32 billion parameters, while DeepSeek's and Alibaba's top open models pass a trillion.
This week: Qwen4-Max is API-only
Alibaba released Qwen4-Max on November 16, 2026, through its API only, at $1.50 per million input tokens and $4.50 per million output tokens. There are no weights to download.
Alibaba's launch post says smaller Qwen4 models are coming to open source, with no date, names or sizes. If they follow Qwen3.8, that will mean open weights under Apache 2.0 or Alibaba's own license, not training code and data. Check each model's license file on Hugging Face when they arrive.
How to tell what a model really offers
- Open the license file. Apache 2.0 or MIT means commercial use with almost no conditions. A custom license means checking for user caps, revenue thresholds and usage policies.
- Look for training code. Inference code lets you run the model. Training code lets you reproduce it.
- Look for the data. A named, downloadable dataset or a detailed data card marks a fully open model.
For most teams that want to run a model privately or fine-tune it, open weights under Apache 2.0 or MIT is enough. Full openness matters when you need to audit what a model learned or prove where its data came from.
Frequently asked questions
What is the difference between open weights and open source AI?
Open weights means the trained parameters are published so you can run and fine-tune the model. Open source AI, under the OSI's definition, also requires the full training code and enough information about the training data to rebuild the model, all under terms that allow any use.
Is Llama open source?
Llama is open weights. Its community license requires companies with more than 700 million monthly active users to ask Meta for a license and adds naming and usage rules, and the OSI says Llama 2 does not pass its Open Source AI Definition.
Which Chinese AI models are open source?
DeepSeek publishes its V4 models' weights under the MIT license, and Alibaba publishes small Qwen3.8 models under Apache 2.0 and its largest Qwen3.8 model under its own license. These are open weights: neither company publishes the training data.
Can I use open weight models commercially?
It depends on the license. Apache 2.0 and MIT allow commercial use with almost no conditions, while custom licenses such as Llama 4's and the Qwen3.8-Max License add thresholds based on users or revenue.
Sources
- The Open Source AI Definition 1.0Open Source Initiative
- Open Source AI Definition FAQOpen Source Initiative
- Open WeightsOpen Source Initiative
- The Open Source Initiative announces the release of the industry's first Open Source AI DefinitionOpen Source Initiative
- Llama 4 Community License AgreementMeta
- DeepSeek-V4-Pro model cardDeepSeek
- DeepSeek-R1 model cardDeepSeek
- gpt-oss-120b model cardOpenAI
- Qwen3.8-27B model cardQwen
- Qwen3.8-2.4T-A95B model cardQwen
- Qwen3.8-Max LicenseQwen
- OlmoAi2
- Olmo-3.1-32B-Think model cardAi2
- Qwen4-Max model pageAlibaba Cloud
- Qwen4-Max: Your New Teammate Has ArrivedQwen


