SenseNova V6

What it is

SenseNova V6 (日日新 V6) is a multimodal foundation model that SenseTime unveiled on 10 April 2025 ^[1]. It is a native mixture-of-experts model with more than 600 billion parameters ^[2]. Mixture of experts means the network contains many specialised sub-networks — the experts — and a routing mechanism sends each piece of input to only a few of them, so the total parameter count can be very large while the computation spent on any single request stays comparatively small. Native multimodal means the model was trained on text, images, audio and video together, rather than having a vision module bolted onto an existing language model.

The three variants

The release covers three models. SenseNova V6 is the general model; V6 Reasoner is aimed at reasoning; and V6 Omni is a lightweight real-time full-modal interactive model that SenseTime called China's first commercialised one of its kind ^[2]. SenseTime said V6 Omni delivers the most advanced multimodal interactive capabilities in China ^[1].

Training on long chains of thought

SenseTime trained the model on over 200 billion tokens of high-quality multimodal long chain-of-thought data, and it supports chains of thought of up to 64K tokens ^[2]. A chain of thought is the intermediate reasoning a model writes out — steps, checks, sub-conclusions — before giving a final answer; letting it run longer generally helps on multi-step problems but costs more computation and memory. Doing this across modalities rather than only in text is the part SenseTime emphasises.

What it can do

The model can analyse full-frame-rate 10-minute videos, compressing a 10-minute video into 16K tokens ^[2]. Compression matters because a model's context window — the amount of input it can hold at once — is finite; turning thousands of frames into a compact token sequence is what makes long video practical to reason over.

On benchmarks, SenseTime said its multimodal reasoning ranked first in China when compared against OpenAI o1, and that it outperformed GPT-4o in data analysis ^[2]. These are the company's own comparisons; independent results were not part of the announcement.

Cost

SenseTime claimed the lowest training and reasoning costs in the industry, and said it had aligned multimodal training efficiency with that of language models ^[2]. Multimodal training is normally more expensive than text-only training because video and audio carry far more raw data for every second of content, so the claim is essentially about making the extra modalities cost about as much as text.

Availability and ecosystem

A new SenseChat app built on V6 entered preview, and V6 could be tried on the SenseChat web platform ^[2]. SenseTime also announced RMB 100 million in subsidies for emerging fields such as embodied intelligence and AIGC ^[2] — money aimed at teams building on top of the model rather than at the model itself.

Where it sits

SenseTime said the unified model showed significant improvements from SenseNova 5.5 to V6 and V6 Reasoner ^[2], placing V6 at the end of a line that had been moving from separate models toward one system handling several modalities and both fast interaction and slower reasoning. Whether the cost and benchmark claims hold up will be settled by outside evaluation and by what developers build with it; the announcement itself establishes the design, the variants and the stated targets.