Table of Contents
An open-weight model lets you download pretrained artificial intelligence and run it on your own computer. While chatbots receive the most attention, image and video generators also offer open-weight options, giving users more autonomy over how they utilize AI.
Understanding Open-Weight Models
In a chatbot, the AI model is the software that generates answers by learning patterns during training and adjusting billions of numerical values known as neural network weights. When a user asks a question, the model uses prompts and those learned patterns to predict the next piece of text. Developers of closed models, such as ChatGPT and Claude, keep their weights private, whereas open-weight models make trained weights publicly available for download.
Pierre Stock, vice president of science at AI company Mistral, told CNET that once a model is open-weighted, the weights spread across servers like Hugging Face, making it virtually impossible to cut off access to it. Using a free model and app allows users to chat locally without subscriptions or usage fees, while also keeping prompts and files off external provider servers.
Users do not need to write code or use a terminal, as desktop applications like LM Studio, Ollama, and Jan enable simple installations that provide a ChatGPT-like interface on a Mac or PC. Users only need a compatible computer, after which they can chat offline. Furthermore, users can adapt models through further training where permitted by their license.
Open-Weight Versus Open-Source
People often use open-weight and open-source interchangeably, but they have distinct meanings. An open-weight model can be open-source, but publishing weights alone does not meet the Open Source Initiative definition. Developers must share the code used to train and run the model, provide sufficient information about training data for a skilled person to replicate it, and allow people to use, study, change, and share the software.
The Open Source Initiative does not require developers to share training data, whereas the Free Software Foundation argues that users should be free to use and change the data themselves. Additionally, some open-weight licenses restrict business use, meaning a model might be free to download but prohibited for commercial applications, which would disqualify it under open-source definitions.
Available Models and Limitations
Independent testing company Artificial Analysis ranks Chinese models high in open-weight intelligence, including DeepSeek, Moonshot AI’s Kimi K3, and Alibaba’s Qwen3.8 family. US alternatives include OpenAI’s gpt-oss models, Google’s Gemma models, Meta’s Llama, Muse Glimmer, and Muse Spark Open, as well as Nvidia’s Nemotron. Startups like Reflection and Thinking Machines Lab offer Beam and Inkling, respectively, while France’s Mistral AI provides Mistral Large 3, the Ministral 3 family, and Mistral Large 4, known as Le Chonk.
Hardware requirements vary, as some larger models require server hardware, while smaller versions operate on personal computers or phones. Open-weight models can still hallucinate, carry training biases, and have private training data that prevents outsiders from fully inspecting their contents. Performance varies by task, and the freedom to modify a model allows users to weaken or remove safeguards.
Stock explained that guardrails are necessary, noting that individuals can partially overcome them with significant effort, akin to fine-tuning a model. Once downloaded, original developers have limited control, though independent researchers can examine models more closely to identify faults. Microsoft and other companies defended open weights in an open letter, arguing that availability helps security teams defend against attackers using AI, as Stock noted that open weighting enables the entire ecosystem to progress together.
Deciding If an Open-Weight Model Is Right for You
Open-weight models draft emails, summarize documents, analyze local files, and assist with coding, but local speed and document handling depend on hardware memory and processing power. Larger models run smoothly on computers with substantial memory or a powerful GPU, such as a Mac with 16GB or 38GB of RAM, while smaller models may run on an 8GB Mac. Buyers must weigh hardware upgrade costs against standard subscription fees for cloud models.
Users can also utilize open-weight models through developer websites like DeepSeek and Kimi, though prompts go directly to the provider. Open weights offer users the ability to retain sensitive documents on their devices, maintain access without internet connectivity, and keep a favorite model even after developers retire online versions, with smaller models increasingly fitting onto devices like smartwatches.
Stock stated that nobody wants to live in a world where access to their model can be abruptly cut off.
