Qianli puts a 10 TOPS NPU inside a 4D radar and hands the chain to models

Yin Qi and Qianli Technology, the company he leads, have put AI inside a millimeter-wave radar. QbitAI's October 27 report describes a 4D radar with a built-in 10 TOPS NPU that supports on-sensor AI perception, a satellite architecture, and both in parallel — with the whole chain from antenna design to point-cloud output handed to models, and perception upgradable over the air.

Traditional 3D millimeter-wave radar is good at range and speed but blind to height. When an overhead gantry and a stopped car in the lane return echoes of similar strength, the sensor cannot tell which is higher, so the system either brakes hard for the gantry or ignores the parked car as an overhead structure. Adding elevation gives range, speed, azimuth and height, turning a flat point cloud into a three-dimensional one.

Adding the dimension is only the first step. More channels and more information bring clutter from guardrails, tunnels, bridges, the road surface and the sky. Rule-based pipelines with fixed thresholds tend to fail in four ways: multipath ghosts, clutter interference, weak-target loss and poor separation of multiple targets. The datasheet looks fine while the road still produces false alarms.

Qianli's approach is to bring algorithms in at the antenna array and waveform design stage, then replace threshold-driven steps — point-cloud detection, angle estimation, enhancement, tracking and classification — with data-driven models. The company's comparison against its previous DSP-based 4D radar: 10x point-cloud density, 40% more detection range, ghost removal up to 95%, and 12x better spatial angular resolution.

The most underrated part is the three operating modes. In edge mode, the radar runs the full chain from ADC sampling to tracking and classification on its own NPU — AI detection, AI-assisted angle estimation, AI point-cloud enhancement — and outputs structured perception results. In satellite mode it concentrates on front-end capture and high-speed transport, handing near-raw data to the central domain controller so larger models can do multi-sensor fusion. In compatibility mode it does both at once.

Running both is not redundancy. A central model wants unprocessed raw returns for early fusion, which requires millisecond alignment and as much signal detail as possible, so its path is inherently longer; braking and warning need an answer now, and fast beats complete. The two timescales cannot be serialized, so they run in parallel. That is a different argument from the industry's older talk of dual-mode redundancy, which is about one path covering for a failure rather than two paths doing different jobs in normal operation.

One concrete example from the report: stable detection of a stationary child at 100 to 150 meters, which buys a longer reaction window for automatic emergency braking in jaywalking and “ghost probe” scenarios. The backdrop is data volume. A 4D point cloud jumps from dozens of points per second to tens or even hundreds of thousands; in satellite mode the raw input can be roughly 100x that of a conventional 4D radar, with ADC bandwidth at 6Gbps. Mainstream assisted-driving cars now carry more than 30 sensors and generate over 10TB a day.

Behind Qianli sit StepFun, where Yin Qi is chairman, and Geely's vehicle programs. The report frames the strategy as “AI plus physical terminals”: start from what the system must accomplish in the physical world and work backwards to what information the sensor should deliver, rather than letting human conventions define what the radar looks for. Yin's earlier line — “brilliance that cannot close the loop is temporary” — sets the real test: whether this radar reaches mass production, and whether over-the-air updates genuinely keep improving perception.