SenseTime releases SenseNova 4.0

Background

Competition among large language model providers had moved beyond answer quality alone. Two other capabilities were becoming selling points: how much text a model could hold in view at once, and whether it could reach outside itself to call tools and other software.

What was announced

On 8 February 2024 SenseTime announced SenseNova 4.0 ^[1]. The main component was SenseChat V4, the conversational model, offered in three versions with context windows of 4K, 32K and 128K ^[1]. A context window is the amount of text — the prompt plus the conversation so far — that a model can take as input in a single pass; a 4K window covers only a short exchange, while a 128K window can take in a long report or a large code file. Offering several sizes lets a user trade cost and speed against how much material the model can read at once, and lets an application pick a smaller, cheaper setting for short chats and a larger one for document work.

The release also added specialised models for data analysis, medicine and vision ^[1]. These are narrower systems built for particular kinds of input or task rather than for open-ended conversation: a model aimed at medical material is not expected to answer general chit-chat, and one aimed at images works on visual rather than purely textual input.

On the developer side, SenseTime launched the SenseChat Function Call & Assistants API ^[1]. Function calling means the model, instead of only producing prose, can decide that a request needs an outside tool — a calculator, a database query, an image processor — and emit a structured request for it; the multimodal part indicates that such calls can involve more than text. An assistants-style API wraps this in a service that keeps track of a task in progress, so that a program can hand over a goal and let the model chain several steps and tool calls together.

SenseTime also presented Office Raccoon, a no-code data analysis assistant, alongside the release ^[1]. No-code means the user works through natural language or a graphical interface instead of writing scripts: the assistant is aimed at people who have data and questions about it but not necessarily programming skills.

Benchmark result

SenseTime reported that SenseChat V4 scored 75.6% on HumanEval, against 74.4% for GPT-4 ^[1]. HumanEval is a benchmark for code generation: the model is given a function signature and a description and must write a program that passes hidden tests. It measures programming ability on a set of short problems, not general competence, so the comparison speaks to one skill rather than to overall quality.

Why it matters

The release is a compact picture of where chat models were heading: several context sizes under one name, specialised sibling models for particular domains, an API for tool use, and a public benchmark comparison against the leading model of the day. Publishing a specific task score next to GPT-4's was a common way for a vendor to argue that its model had reached a comparable level on at least that task, and to give developers a concrete reason to try the API and the assistant product alongside it.