On October 4, Axios reported that Nvidia-backed AI company Reflection is preparing to release its first model. A company spokesperson declined to comment, however. There is no date, model name or download yet.
Reflection says it intends to publish model weights, research and open software around its models. Weights are the huge collections of numbers a model learns during training. When they can be downloaded, developers can often run and adapt the model with more control of their own.
But weights are not the whole recipe. They do not automatically reveal how the training material was selected or how the training was carried out. The Open Source Initiative also requires enough information about the data and the complete code for training and running the system before AI can be called fully open source.
When the model is released, there are a few simple questions to ask: What can actually be downloaded? Which licence applies? Are the data and training method described? What kind of computer is required?
Until then, Reflection's announcement is an interesting direction, not a finished tool or proof of performance.