Qianli Technology, under Yin Qi, releases a 4D millimeter-wave radar with a built-in 10 TOPS NPU

A 4D millimeter-wave radar with a built-in 10 TOPS NPU, supporting local AI perception, a satellite architecture, and multiple output modes in which both operate in parallel.

In plain terms, this radar hands the entire chain—from antenna design to point cloud output—over to models. Borrowing a phrase from the intelligent driving industry, it has in a sense become "end-to-end," and can even continuously upgrade its perception capabilities via OTA.

Familiar technology, familiar playbook—but put them together, why does it feel so unfamiliar?

3nm chips with hundreds to over a thousand TOPS are already on vehicles; does a millimeter-wave radar really need its own independent "brain"?

On closer inspection, this is not some edge attempt by a small company—it is the latest progress from Yin Qi and Qianli Technology, which he helms.

Beyond the mainstream routes of pure vision, vertical integration, and full-stack in-house development, Qianli Technology is more like the chain-style collaboration of a platform company.

That makes it all the more worth examining—what exactly is going on here?

Imagine a scenario: you are driving, and ahead overhead is a gantry, while a car is stopped in your lane. A traditional millimeter-wave radar receives two echoes at the same time, with similar strength, but it cannot tell which is higher and which is lower.

The result is either slamming on the brakes because it treats the gantry as an obstacle, or ignoring the stopped car as an overhead structure. This is the old flaw of the most traditional 3D millimeter-wave radar: it is good at ranging and speed measurement, but lacks the "up-down" dimension.

4D radar fills in exactly that dimension. With elevation angle added, the radar can output information across four dimensions—range, speed, azimuth, and height—and the point cloud spreads from a plane into three dimensions.

Traditional approaches that rely on fixed thresholds and manually written rules to make judgments tend to run into four types of problems in such environments: multipath ghosting, clutter interference, loss of weak targets, and insufficient separation of multiple targets. The sensor datasheet may look impressive, but on real roads it still raises false alarms when it should not and still loses targets when it should not.

The hardware potential of 4D radar has long been stuck at the processing stage, unable to be released.

Qianli Technology's solution is to bring algorithms in from the antenna array and waveform design stage, and then replace key steps such as point cloud detection, angle measurement, point cloud enhancement, and target tracking and classification with data-driven models, reducing reliance on manual rules and empirical thresholds.

Compared with Qianli's previous-generation 4D radar, which used DSP-based on-device computing, point cloud density is enhanced 10-fold, detection range is increased by 40%, ghost elimination is improved 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-side mode, the echo completes the entire chain from ADC sampling to target tracking and classification at the radar end, with models such as AI Detector, AI-assisted angle measurement, and AI point cloud enhancement deployed on the NPU, and the radar directly outputs structured perception results. The advantage is a short local chain and fast response, reducing the computing and communication burden on the domain controller, 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, so the central platform can use larger models to perform earlier-stage spatiotemporal alignment and global reasoning across multi-source information such as cameras and millimeter-wave radar.

The third is compatibility mode, doing both jobs at the same time.

The radar's most raw 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 pass of processing.

Why must it be done twice at the same time? Because these two tasks operate on different time scales.

The central model wants unprocessed raw echoes; to perform multi-source early fusion, it needs millisecond-level alignment precision and signal detail as complete as possible, which is naturally slower and involves a longer chain. But the execution layer on the vehicle cannot wait for that rhythm—braking and warnings need conclusions usable right now, and speed matters more than completeness.

One needs to "see as fully as possible," the other needs to "deliver immediately"; these two requirements cannot be strung together in series, so they can only run in parallel. In the past, the industry's talk of dual modes usually came down to the phrase "safety redundancy," meaning that if something goes wrong with the central system, the radar itself can cover for it.

But redundancy is about stepping in during a failure, whereas the two time scales are about each side having its own job under normal conditions—the latter is the real weight of this design.

The starting point of the change is that the requirements of intelligent driving themselves have changed.

As L2 evolves toward L3 and L4, scenarios such as rain and fog, backlight, nighttime, strong reflections, crossing traffic, and stationary obstacles require perception not only to "see" but also to "see clearly, judge accurately, and react quickly."

To satisfy these scenarios, multi-sensor early fusion has become an unavoidable path. What it needs is not just more sensitive sensors, but a data form that preserves more raw information and can be understood uniformly by models. This brings new problems.

4D radar point cloud density has jumped from a few dozen points per second with traditional radar to tens of thousands or even one hundred thousand points per second; under the satellite architecture, the raw data input volume may be about 100 times that of traditional 4D radar, with ADC bandwidth reaching 6 Gbps.

Mainstream intelligent driving models have more than 30 sensors, generating over 10 TB of data per day; the central computing platform faces just such a river of surging traffic.

Therefore, even just to reduce the output latency of intelligent driving models and efficiently utilize central computing power, specialized tasks such as data preprocessing need to be handed to specialists rather than pushed entirely to the back end.

It is no longer enough for radar to answer only "what target was detected" as before; it must also answer "in what form, and how much effective information, should be handed to the back-end model."

Some might argue: haven't radar products always had on-device data processing modules?

The difference lies in who writes the rules. Traditional 4D radar relies on fixed thresholds and manually written rules to make judgments; 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 base cannot provide an answer, and the system can only handle it in the most conservative way—either braking falsely or missing it. Every time a new type of situation is encountered, it must be redefined, recalibrated, and revalidated, a cycle that takes several months.

After AI adoption, the nature of this changes. The model learns patterns from data, and when it encounters combinations that did not appear in training, it can still make judgments based on learned features without waiting for a new rule to be written. For a long-tail problem, the cycle from collection, feedback, training to release is compressed from months to weeks or even shorter.

The first thing Qianli Technology did was put AI computing inside the radar, letting the sensor process the data it collects to a deeper level first and directly output structured results. The second thing was to decide how to deliver the results after computing—that is, the three modes mentioned earlier: how much of the task stays in the sensor and how much data is handed to the central unit, configured according to platform and function level.

The distinction among the three architectures can be viewed along three lines: function level, EE architecture, and cost and control rights.

Target-level output anchors conventional assisted driving and distributed EE architectures and is cost-sensitive; the satellite architecture starts from L2+ and becomes a rigid choice under the front-fusion requirements of L3/L4, because only by sending raw data to the central unit can millimeter-wave echoes have the opportunity to perform feature-level fusion with image features in a unified BEV space.

The real engineering boundary lies in bandwidth and platform capability: if domain controller computing power and high-speed interfaces are insufficient, the satellite architecture loses its physical basis for implementation; if the function level stays at L2 and there is no real demand for front fusion, paying transmission and computing costs for raw data does not hold up on the mass-production ledger.

This radar from Qianli retains all three paths at the same time, precisely corresponding to the transition period in which the vehicle platform moves from distributed to central centralized, where the same set of hardware must find its place across different models and different function levels.

Compatibility mode is rather special in this analysis because it occupies two lines at once: for platforms that have not yet completed centralization, the edge-side result can be used directly; for platforms that already have domain controller computing power, RAW data can enter the central unit for front fusion. What is most valuable during the transition period is precisely this "can connect at both ends."

For example, the Qianli AI radar's stable detection distance for a stationary child is 100 to 150 meters, meaning that in scenarios of a child crossing or a sudden pedestrian appearance, it can provide a longer reaction window for AEB.

AI post-processing and multipath feature modeling eliminate about 95% of ghosting and false positives, and 4D point cloud enhancement can also assist in judging passable space.

Going deeper into software-hardware integration, the research and development of the entire sensor has also changed: from waveguide antennas, RF front ends, SoC/NPU, and modulation methods to models and the data closed loop, all are designed collaboratively. On the surface it is a hardware upgrade; in essence, the sensor shifts from a "target-level reporter" to a "feature-level collaborator."

In the early years, Yin Qi already 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 form was deliberately made to lean toward a combination of software and hardware.

But the dilemma of AI 1.0 was that it had algorithmic depth but lacked a terminal carrier capable of scale and continuously generating real feedback, so the closed loop was never truly closed. That is why he later said, "Any glory that cannot be closed-loop is temporary."

The combination of Qianli and Step supplements the side that the creed had always lacked. The key lies in the homologous input for pretraining of the "native foundation model": general corpora are fed in together with real intelligent driving perception and planning-control data, so the foundation model is born with cognitive structures for three-dimensional space and time series.

Yin Qi heads both companies at the same time, filling the gap commonly found between model companies and terminal companies; the cycle in which "the brain defines what the terminal outputs, and the terminal feeds back what the model learns" has started turning under the same helmsman.

Starting from the iteration of the model paradigm to change the underlying sensors and computing architecture—this is the most noteworthy part of Qianli's attempt this time.

In the past it was "what humans want radar to see"; now it works backward from the perspective of AI models and Agents: what tasks this system must complete in the physical world, what form of information it needs, how far the effective distance must be, and how high the resolution capability must be.

This fundamentally shields against interference from the "human perspective" and "paradigm inertia."