A few days ago, Google ushered in the era of pocket AI with Gemini Nano, a generative text AI model that can work directly from a mobile phone and without a cloud connection. Now Microsoft has surprised everyone by releasing its competitor, the Small Language Model (SLM), called Phi-2, which can also be used on mobile phones.

Phi-2. at Microsoft Research they stated This new text-to-text model is stated to be “small enough to run on a laptop or mobile device.” Satya Nadella already told us about this at the recent Ignite 2023 event, and this is the evolution of Phi-1, which has only 1.3D parameters.

A ChatGPT in a compact format. We are looking at a model with 2.7 billion parameters (2.7D); This figure is a figure that firmly guides its use in more modest devices. In Google Gemini Nano, for example, there are two versions with parameters 1,800 (1.8B) and 3.25 billion (3.25B).

Tiny but tough. At least according to those responsible. Reason? It is capable of equaling or even exceeding the performance and accuracy of theoretically more capable models such as the Llama 2-7B or Mistral-7B. In tests carried out internally, performance is excellent, despite having almost a third of the parameters of these models.

It also surpasses the Gemini Nano. Microsoft’s new model is also capable of surpassing the performance of Google’s most ambitious “mobile” model, the Gemini Nano 3.25B, and according to the company’s researchers, it is capable of offering less “toxicity” and more subjective responses based on the answers given. for example by Lama 2.

Comparisons are hateful. At the presentation of the model, Microsoft reminded that the spectacular Gemini Ultra show was not that spectacular. He cheated a little. In one of the episodes, Google showed how Gemini Ultra could help a student with physics problems, and at Microsoft they used the same prompts as Phi-2 and it appeared to be able to solve these doubts as well. A much smaller model..

Just to research. Despite this capacity, Phi-2 will not be able to reach our mobile phones, at least for now. The model is licensed for “research purposes” only and cannot be used commercially. The possibility of using it for “non-revenue-generating, non-commercial purposes” is mentioned, but its practical application is a little less direct. However, this can be accessed through the catalog of the new Azure AI Studio platform.

in Xataka | ChatGPT is getting lazy and OpenAI knows it. But if you give him a tip (or tell him you don’t have fingers) he’ll do better.