Open Weights

Open weights differ from fully open source: the former makes model parameters public but does not necessarily disclose training data and complete training code.

开放权重模型指参数权重文件向公众开放下载，可本地部署微调，但训练数据和完整源码不一定公开，区别于完全开源

Meta adopts an open-weights strategy, making model parameters publicly available for research and commercial use, though the training data and full pipeline are not entirely transparent.

Open-weight models refer to models that, after training, publicly release the neural network parameter weights under an open license, allowing them to be downloaded and run on one's own hardware and fine-tuned with proprietary data without paying inference fees.

An open-weights strategy means that model parameters can be downloaded and used for free, but training data and the complete training process are generally not made public, and it has become the mainstream choice for companies such as Meta Llama and Mistral.

开放权重模型仅发布训练完成的模型参数文件，而严格意义的开源AI要求公开训练代码、数据集与完整复现流程

开放权重（Open Weight）仅指模型的参数权重文件可供下载和使用，但训练数据、完整训练代码、RLHF标注数据等核心资产往往并不公开，与真正意义上的开源（Open Source）存在区别

Open-weight models let users download the parameters, run them on their own servers, fine-tune them and redistribute them; hosted services provide only an API, and users cannot access the model itself.

In recent years, open weights have become a key policy topic in the AI competition between the United States and China.

Meta's continued release of open-weight models lowers the barrier to AI application development, while potentially squeezing the survival space of truly open-source projects.

Some AI projects adopt an open-weights model that publishes model parameters but restricts commercial use, and such practices often spark controversy over the definition of Open.

The author believes open weights are crucial to FollowUp Local, for reasons including: chat records contain customer privacy and local running keeps data from leaving the computer; local inference has zero marginal cost, making a one-time purchase model feasible; the 1B or 4B model can be freely chosen and fine-tuned on industry data; and even if the author stops maintaining it, the tool and model can still keep running on the user's machine.

The reasoning chains of closed-source models are often returned in encrypted or summarized form, making it impossible for outsiders to verify whether logical shortcuts or hidden biases exist; open weights, in contrast, allow the reasoning process to be reproduced and audited by third parties, thereby turning safety from a "vendor promise" into something "community-verifiable".

Open weights mean the model parameters can be downloaded, inspected, modified and redistributed, whereas a closed-source API exposes only an inference interface.

With model weights in hand, researchers can do things an API cannot: view intermediate-layer representations, continue training or fine-tune, deploy offline, and reproduce and audit results.

The value of open weights is often realized in scenarios the publisher did not anticipate.

A compromise of "open weights but restricted capabilities" has emerged in the industry, such as delayed release, tiered licensing, or capability-threshold controls.

The key consequence of open weights is that once the weights are released, they cannot be remotely shut down or monitored in real time like a cloud API, and users can freely deploy, modify and redistribute them in a completely offline environment.

The real benefits brought by open weights include: auditability (researchers can inspect model behavior), customizability (enterprises can fine-tune locally to meet needs such as keeping data within their domain), and ecosystem competition (exerting pressure on closed-source vendors, curbing price monopolies and driving capability improvements).

Open weight means the model parameters can be downloaded and executed locally, but the training code, training data, and terms of use are not necessarily open; Gemma uses a custom Gemma Terms of Use.