SenseNova 4.0

What it is

SenseNova 4.0 (日日新 4.0) is an upgrade of SenseTime's foundation model sets, announced on 8 February 2024 ^[1]. Its core model is SenseChat V4, and the release also covered a group of specialised models for data analysis, medicine, vision and image generation ^[1].

Origins

The 2023 SenseNova set was the predecessor of this release. SenseTime had said by February 2024 that more than 3,000 enterprise users had adopted SenseNova since April 2023 ^[1]. SenseNova 4.0 was the company's next step from that first set.

Key features

SenseChat V4 supports context windows of 4K, 32K and 128K tokens ^[1]. A context window is the amount of text, measured in tokens, that a model can take into account in one exchange. Longer windows let a model work with longer documents or conversations.

The family included SenseChat V4, SenseChat-DataAnalysis V4, SenseChat-Medical V4, the 30-billion-parameter multimodal SenseChat-Vision V4 and the 10-billion-parameter text-to-image model SenseMirage V4 ^[1]. Multimodal means the model works with more than one kind of input, such as text and images.

SenseTime introduced its first multimodal tool-calling API, the SenseChat Function Call & Assistants API. It connects models to search, code interpretation, image question answering and text-to-image tools ^[1]. Tool calling lets a model ask an outside program to do a job, such as running code, and then use the result in its answer.

The release also featured Office Raccoon, a no-code data analysis tool ^[1]. A no-code tool lets users analyse data without writing programs.

Results

SenseTime's report gave SenseChat V4 a 75.6% one-pass rate on HumanEval, a set of code-writing problems, against 74.4% for GPT-4 ^[1]. HumanEval results are the company's own test figures, so they show how the model did on that benchmark rather than in general use.

On the MME benchmark for vision models, SenseChat-Vision V4 scored 2199.5, against 1926.57 for GPT-4 ^[1]. SenseChat-DataAnalysis V4 reached 85.71% accuracy on more than 1,000 questions ^[1]. SenseMirage-Turbo V4 was reported to run inference 10 times faster than the V4 base model ^[1]. Inference is the step in which a trained model produces an answer.

Position in development

SenseNova 4.0 sits between the first SenseNova set of 2023 and the 5.0 release that followed in 2024. It moved the series from a set of products into a family of specialised models with tool calling. The comparisons in the report use GPT-4 as the reference point.