Only a couple of months ago, it was revealed Called the 3.1 to 405 billion parameters. Bugs introduced yesterday, the new model Called 3.2, which you can analyse and understand tables, graphs, and images. Offered in four variants, two of which are optimizuara for the functionality on the mobile device. Of course, they're still open source. Developers and businesses can use them, too, through the Amazon Bedrock, Google Cloud Vertex HIM, and Microsoft Azure to HIM.
Spot Called 3.2 in the us, Europe
The pattern Called 3.2 in the case of small and medium-sized enterprises (respectively, 11 and 90 billion parameters) provide properties that are more advanced, and the performance is higher. It can analyze and understand the contents of the documents with graphs and tables, and to create, to paint, including the titrat), and to modify the image.
According to the tests, to be issued from the Spot, Called 3.2 11B and 90B in excess of Claude 3 of the Haiku, and the GPT-4o mini within the meaning of the text and the image. Support multimodale call for the architecture and training of the new sets of data up-to-date. These two models are not available in Europe, as indicated by the company's Menlo Park, since it is not possible to use the user's data (Meta, it requires greater clarity and regulatory framework).
The pattern Called 3.2-1B, and 3B, with the 1 and the 3 billion parameters can be executed on the mobile device, so without access to the cloud. This improves performance and ensures privacy. They can be used, for example, to rewrite the text, in order to create drafts of and for the generated summaries (they do not have multimodale).
They are optimizuara for the device to the Arm. Meta collaboration with Qualcomm and MediaTek in the process. The following models are not trained on the data from the user's european, so that they can be used in Europe from the manufacturer and the developer of the application for a smartphone and a tablet.
Discussion about this post