SenseTime launches SenseNova 5.0, benchmarking against GPT-4 Turbo

Background

SenseNova 4.0 had been released in February 2024. The 5.0 release, two months later, compared the new model against GPT-4 Turbo, a model from OpenAI.

What happened

SenseTime released SenseNova 5.0 on 23 April 2024 at Tech Day events in Shanghai and Shenzhen ^[1]. The company said the model matched or beat GPT-4 Turbo on mainstream objective benchmarks ^[1].

Key details

The mixture-of-experts (MoE) model was trained on more than 10TB of tokens and supports about 200K tokens of context ^[1]. A token is a unit of text a model reads and writes, so the context figure describes how much text the model can take into account at once.

The launch also introduced edge-side models and an enterprise edge device for finance, coding, healthcare and government ^[1]. Edge-side means the model runs on a device rather than in the cloud. Xu Li, the chairman and CEO, said that SenseTime aimed to push the boundaries of large-model capabilities ^[1].

Shares in SenseTime surged more than 30% after the launch and trading was suspended ^[2].

Impact

The launch put the company's product claims and its share price movement in the same news coverage, and for readers the event shows both together.

Terms used here

A mixture-of-experts model splits its network into many specialised parts, called experts, and sends each input to only some of them. This keeps the work for each answer lower than in a model that uses every parameter every time, although the full model still has to be stored and served.

Edge-side deployment means running a model on a device such as a phone or car system rather than only in the cloud. Benchmark comparisons show results on chosen tests, not general use. A share price is what one share trades for in the market, and a jump on the day of an announcement reflects that day's trading rather than an independent measure of the product.