SenseTime unveils SenseNova 5.5 and real-time SenseNova 5o at WAIC

Background

The World Artificial Intelligence Conference (WAIC) is held in Shanghai. SenseTime had released SenseNova 5.0 two months before the 5.5 event, and it used WAIC to show the next step in its line.

What happened

SenseTime unveiled SenseNova 5.5 at WAIC in Shanghai in July 2024 ^[1]. It presented SenseNova 5o as China's first real-time multimodal model, comparable to GPT-4o's streaming interaction ^[1].

Key details

SenseTime said performance improved 30% over SenseNova 5.0 ^[1]. It claimed that SenseNova 5.5 beat GPT-4o on 5 of 8 key metrics ^[1]. Those are the company's own results.

A free 50-million-token package was offered to enterprises moving from OpenAI ^[1]. Token packages are a common way for model providers to offer trial usage to customers.

Xu Li said that 2024 was a critical year as large models evolved from unimodal to multimodal ^[1].

Impact

The event presented SenseTime's real-time multimodal model together with a free offer to enterprises leaving OpenAI, so the technical claims and the customer offer were presented at the same event.

Terms used here

A real-time multimodal model can take speech, images and text as input and respond while the user is still interacting. Streaming means the output is sent in pieces as it is generated, rather than all at once at the end. A lightweight model is a smaller version of a larger model, designed for faster responses and lower computing cost.

An API (application programming interface) is the set of rules that lets other software send requests to a model. Moving from one provider's API to another usually means changing how requests are sent and checking that the answers still behave as expected, which is what migration consulting covers. Free token packages let a customer test a model at no cost before paying for regular usage. Benchmark claims compare models on selected tests, so the five-of-eight figure is best read as the company's own count.