SenseNova 5.5

What it is

SenseNova 5.5 is a release of SenseTime's SenseNova model family, unveiled at the World Artificial Intelligence Conference (WAIC) in Shanghai in early July 2024^[1]. It is not a single model but a set of models at different sizes and for different places to run: a main model, a lightweight one built for fast serving, one small enough to sit on a device, and a multimodal model built for real-time conversation^[1].

The multimodal piece is called SenseNova 5o, which SenseTime described as China's first real-time multimodal model, matching GPT-4o's streaming interaction for real-time conversation and speech recognition^[1]. Streaming interaction means input is processed as it arrives rather than uploaded as one finished batch, so a user can speak to the model and hear it answer without waiting for a full recording to be sent first.

Where it came from

The release came roughly two months after SenseNova 5.0^[1]. SenseTime presented the new version as a step up rather than a rebuild: it said overall performance rose 30% over 5.0, with gains in mathematical reasoning, English and instruction following^[1]. Instruction following is the ability to do what a prompt actually asks — the requested format, the stated constraints, a multi-step task — and it is often what separates a demonstration from something usable in production.

The company also claimed that the model beat GPT-4o on 5 of 8 key metrics^[1]. Comparisons of this kind depend heavily on which benchmarks are picked and how they are run, so the figure is best read as the company's own positioning rather than an independent result.

Main features

The lightweight member of the family, SenseChat Lite-5.5, was reported to cut inference time by 40%, down to 0.19 seconds, while reaching 90.2 words per second^[1]. Inference time is the delay between a request going in and an answer coming back, and it is what decides whether a chatbot feels instant or sluggish when many people use it at once.

At the other end of the range, SenseTime offered an edge-side model for as little as RMB 9.90 per device per year^[1]. An edge model runs on the device itself rather than in a data centre, which is how functions can work without a network round trip; pricing it per device per year makes the cost predictable for hardware makers rather than tied to how much text each user generates.

The release also came with a commercial offer. Under a programme called "Project $0 Go", enterprises moving from OpenAI were offered a free 50-million-token package plus API migration consulting^[1]. A token is the small unit of text — roughly a word fragment — that models read and are billed by, so a token package is effectively prepaid usage; the consulting covers the work of rewriting calls to a different API.

Beyond the language models, the release introduced Vimi, a controllable avatar video generator, and upgrades to Raccoon^[1]. SenseTime said more than 3,000 government and corporate customers were using SenseNova^[1].

Where it sits

Later company reporting filled in the timing and the shape of the family. SenseTime's interim results, published on 29 August 2024, said SenseNova 5.5 had been upgraded in July 2024 and had approximately 30% higher overall capability than version 5.0^[2]. The same results said the company had developed several model sizes based on SenseNova 5.5^[2].

Taken together, the release is less a single benchmark claim than an attempt to cover the whole deployment range at once — real-time multimodal interaction at the top, cheap fast serving in the middle, per-device edge pricing at the bottom — and to win over teams already building on a competitor's API.