A 4D millimeter-wave radar with a built-in 10 TOPS NPU, supporting local AI perception, a satellite architecture, and multiple output modes running both in parallel.。
To put it plainly, from antenna design to point cloud output, the entire chain of this radar is handed over to the model. If we borrow a term from the autonomous driving industry, it is in a sense also "end-to-end," and can even continuously upgrade its perception capabilities via OTA.
Familiar technology, familiar tactics—but put together, why does it all feel so unfamiliar?
"3nm chips delivering hundreds to over a thousand TOPS have already made it into cars — is it still necessary for a millimeter-wave radar to have its own independent "brain"?"
A closer look reveals this is no fringe experiment by some small company — it'sYin Qiand the one he helmsQianli TechnologyThe latest developments.
On the path of physical AI, alongside Qianli there are also StepFun, chaired by Yin Qi, as well as Geely Auto's complete-vehicle scenarios.
Beyond the mainstream routes of pure vision, vertical integration, and fully in-house stack development, Qianli Technology is more like the chained collaboration of a platform-type company.
Then we'll have to take a look — what exactly is going on?
What kind of radar did Qianli make?
Imagine a scenario: you are driving, with a gantry overhead ahead and a car parked in your lane. A traditional millimeter-wave radar receives two echoes of similar strength at the same time, but it cannot tell which is higher and which is lower.
As a result, it either treats the gantry as an obstacle and slams on the brakes, or ignores the parked car as an overhead structure. This is the classic ailment of traditional 3D millimeter-wave radar: it is good at measuring distance and speed, but lacks the “up-down” dimension.
4D radar fills in exactly this dimension. Once the elevation angle is added, the radar can output information across four dimensions—distance, velocity, azimuth angle, and height—and the point cloud changes from a flat spread into a three-dimensional one.
The first step of Qianli's AI radar is to fill in this key dimension. But as the number of channels and the amount of information increase, the burden of signal processing grows as well. Strongly reflective objects such as guardrails, tunnels, bridges, the ground, and the sky introduce large amounts of clutter into the echoes.
Traditional solutions that rely on fixed thresholds and manually crafted rules for judgment are prone to four types of problems in such environments: multipath ghosting, clutter interference, loss of weak targets, and insufficient separation of multiple targets. The sensor's spec sheet looks great, but on real roads it still produces false alarms when it shouldn't and still drops targets when it shouldn't.
For a long time, the hardware potential of 4D radar has been stuck at the processing stage, unable to be unleashed.
Qianli Technology's solution is to bring algorithms in at the design stage of the antenna array and waveform, and then replace key stages—point cloud detection, angle measurement, point cloud enhancement, target tracking and classification—with data-driven models, reducing dependence on manual rules and empirical thresholds.
The AI in this radar gets involved from the very front end of signal generation and runs all the way through to the end of environment understanding.
Compared with Qianli's previous-generation 4D radar that used DSP-based edge computing, the point cloud density is enhanced 10-fold, detection range is improved by 40%, ghost elimination is raised to 95%, and spatial angular resolution is improved 12-fold.
Next comes the most easily underestimated part of this radar: it supports three operating modes. In edge mode, the echoes complete the full pipeline on the radar itself, from ADC sampling to target tracking and classification, with AI Detector, AI-assisted angle measurement, AI point cloud enhancement, and other models deployed on the NPU; the radar directly outputs structured perception results. The advantages are a short local pipeline and fast response, reducing the computational and communication burden on the domain controller side, making it suitable for platforms that need mature object-level interfaces or distributed perception capabilities.
In satellite architecture mode, the radar centers on front-end acquisition and high-speed transmission, handing data closer to the raw signal to the ADCU or central domain controller for unified processing. The central platform can use larger models to perform earlier-stage spatiotemporal alignment and global reasoning on multi-source information such as cameras and millimeter-wave radars.
The third is compatibility mode, doing both jobs at the same time.
The radar's rawest RAW data is handed directly to the SOC for processing and fusion, while at the same time the edge-side processor on the radar independently completes a round of processing on its own.
Why process everything twice? Because the two tasks operate on different time scales.
The central model wants unprocessed raw echoes; it aims to do multi-source early fusion, which requires millisecond-level alignment accuracy and signal detail that is as complete as possible—naturally slower, with a longer pipeline. But the vehicle's actuation layer cannot wait for that pace; braking and warning need conclusions that are usable right now, where fast matters more than complete.
One demands “see as completely as possible,” the other demands “deliver immediately.” These two requirements cannot be chained together; they can only run in parallel. In the past, when the industry talked about dual mode, it usually came down to the phrase “safety redundancy,” meaning that if the central system ever failed, the radar itself could hold the line.
But the fallback refers to stepping in when a fault occurs, while the two time scales refer to each side having its own tasks under normal conditions — and it is the latter that carries the real weight of this design.
Behind this lies a natural question: which part of AI computing demands should be placed in the radar, and which should be kept in the central domain controller for processing?
Models — how do they change the radar?
The starting point of the change is that the requirements of intelligent driving itself have changed.
As L2 evolves toward L3 and L4, scenarios such as rain and fog, backlighting, nighttime, strong reflections, cross-traffic, and stationary obstacles require perception to not just "see," but also "see clearly, judge accurately, and react quickly."
To meet these scenarios, multi-sensor early fusion has become an unavoidable path; it requires not just more sensitive sensors, but a form of data that preserves more raw information and can be understood by models in a unified way. This brings a new problem.
4D radar point cloud density has leapt from the few dozen points per second of traditional radar to tens of thousands or even 100,000 points per second; under a satellite architecture, the raw data input volume may be about 100 times higher than that of a traditional 4D radar, with ADC bandwidth reaching 6Gbps.
Mainstream smart-driving models now carry more than 30 sensors and generate over 10TB of data per day, and what the central computing platform faces is a river of surging traffic like this.
So, even if only to reduce the output latency of the intelligent driving model and make efficient use of centralized computing power, it is necessary to hand over specialized tasks like data preprocessing to the professionals, rather than pushing everything to the backend.
It is no longer enough for radar, as before, to only answer "what targets have been detected"; it must also answer "in what form, and how much effective information, can be handed to the back-end model."
Some may argue: haven't past radar products always had on-device data processing modules?
The difference lies in who writes the rules. Traditional 4D radar makes judgments based on fixed thresholds and hand-written rules; the rules are written by humans, and humans can only write down situations they have thought of. For a corner case never seen before, the rule library has no answer, and the system can only respond in the most conservative way—either braking erroneously or missing the object. Every time a new type of situation appears, it must be redefined, recalibrated, and revalidated, a cycle that takes several months.
After the AI transformation, the nature of this matter changed. Models learn patterns from data, and when encountering combinations that never appeared during training, they can still make judgments based on the features they have already learned, without needing to wait for a new rule to be written. The cycle for a long-tail issue—from collection, flow back, entry into training, to release—has been compressed from being counted in months to being counted in weeks or even shorter.
The first thing Qianli Technology did was put AI computing inside the radar, letting the sensor process its own collected data to a deeper level and directly output structured results. The second thing was deciding how to deliver the results after computing—that is, the three modes mentioned earlier: how many tasks to keep in the sensor and how much data to hand over to the center, configured according to the platform and function level.
The distinctions among the three architectures can be examined along three lines: functional hierarchy, EE architecture, and cost versus control.
Target-level output is anchored in conventional assisted driving and distributed EE architectures, which are cost-sensitive; satellite architectures start from L2+ and become a rigid choice under the early-fusion requirements of L3/L4, because only by sending raw data to the central platform do millimeter-wave echoes get a chance to fuse at the feature level with image features in a unified BEV space.
The real engineering boundaries lie in bandwidth and platform capability: if domain controller compute power and high-speed interfaces are insufficient, the satellite architecture loses its physical foundation for implementation; if the functional level remains at L2 and there is no genuine need for pre-fusion, paying transmission and compute costs for raw data does not hold up on a mass-production ledger.
The ability of Qianli's radar to retain three paths simultaneously corresponds precisely to the transition period when the vehicle platform moves from distributed to centralized architectures, where the same set of hardware must find its place across different vehicle models and different functional tiers.
In this analysis, compatibility mode is rather special: it occupies two lines at once. On platforms that have not yet completed centralization, the edge-side result can be used directly; on platforms that already have domain-control compute, RAW data can also go to the center for early fusion. During the transition period, what is most valuable is exactly this "ability to connect at both ends."
For example, the Qianli AI radar's stable detection distance for stationary children is 100 to 150 meters, meaning that in child-crossing and ghost-appear (sudden pedestrian emergence) scenarios, it can provide a longer reaction window for AEB.
AI post-processing and multi-path feature modeling eliminate ghosts and false alarms by about 95%, and 4D point cloud enhancement can also assist in judging passable space.
Yin Qi's Physical AI Closed Loop
The deepest logic behind Qianli Technology's AI radar is that "modal AI" has begun to reshape the underlying sensors.
Going deep into software-hardware integration, the sensor's entire development has changed: from waveguide antennas, RF front ends, SoC/NPU, and modulation schemes to models and the data closed loop, everything is co-designed. On the surface it is a hardware upgrade, but in essence the sensor is shifting from a "target-level reporter" to a "feature-level collaborator."
Radar AI-ization is not optional; the perception end is the entry point of data and intelligence, and millimeter-wave radar cannot be bypassed on this path.
Years ago, Yin Qi took "AI + physical terminals" as a foundational creed: 3D cameras are the bridge connecting the digital and physical worlds, AIoT is "AI + IoT + space," and the organizational structure was deliberately built to lean toward hardware-software integration.
But the predicament of AI 1.0 was that while it had algorithmic depth, it lacked a terminal vehicle that could scale and continuously generate real feedback, so the closed loop never truly closed. That is why he later said, "glory that cannot close the loop is only temporary."
The combination of Qianli and StepFun fills in the side the creed has always lacked. The crux lies in the pretraining same-source inputs of the "native foundation model": general corpora together with real autonomous driving perception and planning-control data are fed in together, so the foundation model is born with a cognitive structure of three-dimensional space and time series.
With Yin Qi at the helm of both companies, the chasm that usually separates a model company from a terminal company is bridged, and the cycle of "the brain defining what the terminal outputs, and the terminal feeding back what the model learns" begins to turn under a single helmsman.
This is the extension of the AI + physical terminal concept in breadth, and it is also a path beyond pure vision, vertical integration, and full-stack in-house development.
For the phrase AI + physical terminals, there are similar expressions in the industry, but most discussion stays at the "AI empowerment" level, holding that AI changes task workflows and provides the value of more automated information presentation.
Starting from the iteration of model paradigms to change underlying sensors and computing architectures — this is the most noteworthy part of Qianli's current attempt.
In the past it was "what humans want the radar to see"; now it is reasoned backward from the perspective of AI models and Agents: what task the system must accomplish in the physical world, what form of information it needs, how far the effective range, and how high the resolution capability.
This blocks, at the root, the interference brought by the "human perspective" and "paradigm inertia."
Digital AI advances on the strength of massive parameters and data that is almost readily available.
But at the present stage, the ceiling of physical AI is determined by the speed of the real-data closed loop and the density of high-quality data. So the very existence of such an attempt already gives Qianli Technology irreplaceable, unique value.