SenseTime drops SenseNova 5.0 and goes big on edge AI

While most AI giants are still obsessing over massive cloud clusters, SenseTime is pushing hard to cram frontier models directly into your phone and office desk.

At its Tech Day event in Shanghai, SenseTime officially unveiled SenseNova 5.0, marking the fifth major iteration of its flagship model family since April 2023. SenseTime Chairman and CEO Dr. Xu Li noted that the company is sticking to scaling laws, structuring the model around a three-tier architecture of Knowledge, Reasoning, and Execution.

MoE architecture and 200k context

SenseNova 5.0 adopts a Mixture of Experts architecture trained on more than 10TB tokens, relying heavily on synthetic data.

During inference, the model supports an effective context window of approximately 200,000 tokens. SenseTime focused its engineering upgrades on mathematics, reasoning, coding, and cross-document analysis to tackle demanding tasks in finance and data analytics.

Multimodal perception also got a major boost. SenseNova 5.0 can digest high-resolution images, handle text-to-image generation, and extract data across multiple documents. It secured the top spot on the MMBench benchmark and posted leading scores on MathVista, AI2D, and ChartQA.

SenseTime also teased upcoming text-to-video capabilities, allowing users to preset character costumes, hairstyles, and scenery to keep visual styles consistent across clips.

Running fast on the edge

Alongside the cloud model, SenseTime launched a full-stack edge-to-cloud product matrix aimed straight at hardware deployment.

Its edge-side language model hits an inference speed of 18.3 words per second on mid-range chips, jumping to an impressive 78.3 words per second on flagship silicon.

Image generation on device is equally snappy. SenseTime's edge diffusion model runs in under 1.5 seconds on mainstream platforms, producing images at 12 megapixels and above with support for outpainting and rotation.

For enterprise clients handling sensitive records in healthcare, government, and finance, SenseTime rolled out an all-in-one edge appliance. The company claims the device slashes inference costs by roughly 80% while utilizing only 50% of CPU capacity during accelerated searches.

Real-world deployments are already live: Kingsoft Office integrated SenseNova into WPS 365, Haitong Securities built vertical compliance and advisory tools on it, and the new Xiaomi SU7 EV tapped SenseTime for its smart cabin.

Why it matters

SenseTime's smartest move here isn't just chasing benchmark bragging rights; it is focusing relentlessly on edge inference and private enterprise hardware.

For businesses terrified of runaway cloud API bills and corporate data leaks, an on-prem appliance that cuts inference costs by 80% is an easy sell. Pushing nearly 80 words per second on flagship mobile silicon opens real doors for responsive, offline AI assistants on smartphones and inside connected vehicles like Xiaomi's SU7.

We think shifting competition from pure cloud parameters to device-level efficiency is the right battleground. The lingering question is whether SenseTime can maintain this efficiency edge as competitors optimize smaller open-weights models for consumer hardware.