Open-weight models: free local runs and licenses

⚡ The gist in 5 seconds - There are now hundreds of open-weight models, but "open" doesn't mean "do whatever you want": each has its own license, and training code and data are usually not published. - Weights are available in safetensors or GGUF formats; nearly everything listed is hosted on Hugging Face. - The main constraint is hardware: two 27-billion-parameter models can differ in memory requirements by several times due to quantization, and your GPU won't handle everything. ### 🔍 What was found The Open Source Initiative, in its Open Source AI Definition (OSAID), identifies four freedoms: to use the system for any purpose, to study it, to modify it, and to distribute it, including modified versions. Open-weight models cover the first and fourth freedoms but not the second: without training code and data, you can't fully study how the model arrived at its weights. That's why the term "open-source neural networks" is imprecise. Among permissive licenses are Apache 2.0 and MIT. Under Apache 2.0 come OpenAI's gpt-oss-120b and gpt-oss-20b (117 and 21 billion parameters, released August 2025), Qwen3.6, Gemma 4, Mistral 7B, Mixtral, and Mistral Small. Google only switched Gemma to Apache 2.0 in April 2026. Under MIT come DeepSeek-V4 and GigaChat3. Conditional licenses: the Llama Community License requires a separate license above 700 million monthly active users, while Kimi K2 uses a Modified MIT with thresholds of 100 million users or $20 million in monthly revenue. MiniMax released M2 and M2.5 under MIT, and M2.7 under a Modified-MIT that is essentially non-commercial: any commercial use requires written permission. ### 💡 Why it matters Before running a model, you should check not only the price but also the license: it's the license that determines whether you can use the model commercially and on what terms. OSAID provides criteria for assessing how open a system is — you can consult them right when choosing a model. A local run makes sense when data can't be sent to someone else's cloud (medicine, finance), when you need offline operation, when there are no provider limits or moderation, or when you need to fine-tune the model for your task. But top closed models still beat open ones on complex tasks, weak hardware cuts speed, and setting up a stack from scratch eats more than one evening. ### 🧩 Context Over the past year, so many open weights have been released that choosing between them has become a task in itself. Open weights are safetensors or GGUF files containing billions of numbers tuned so the model predicts the next token. What was in the training data and how it was selected is generally not disclosed. That's why the OSI introduced OSAID with its four freedoms: open weights without training code and data don't provide the full ability to study and modify the model. The practical guide in the article also answers whether a specific model will fit in your GPU and when a local run is cheaper than a cloud API.

🔗 Read on thecode.media