This week a tutorial surfaced in the Horizon Robotics developer community and on Juejin: it walks through converting YOLOv5x (an open-source visual detection model) from raw PyTorch weights all the way to an HBM model (Horizon's proprietary model format) executable on the Journey J6 chip, completing three steps — ONNX export, quantization calibration, and compilation. This is not a research breakthrough, but it is an industry signal: the toolchain for domestic automotive-grade (automotive component spec) AI chips is closing its final gap.
What this is
Horizon Robotics is a domestic company building automotive-grade AI chips for intelligent driving. Its latest platform, Journey J6, targets end-to-end autonomous driving and robotics scenarios. YOLOv5x is an open-source visual detection model that industry has used for five years, capable of recognizing pedestrians, vehicles, and obstacles in images — commonly deployed as a front-end perception module in autonomous driving.
Getting this model onto the J6 chip is not just installing software. Developers need to convert the PyTorch model to ONNX (a common model interchange format), perform PTQ quantization (compressing floating-point parameters to lower bit widths to fit automotive compute budgets), then use Horizon's in-house compiler to produce an HBM file the chip can execute, and finally write C++ inference code on the board. The tutorial covers the first three toolchain steps; board-side deployment is left for a follow-up.
Industry view
The optimistic camp reads the tutorial as evidence that Horizon's developer ecosystem is becoming self-service — work that only the chip vendor's own engineers used to do can now be completed by ordinary algorithm engineers following documentation. The historical "usability" weak spot of domestic chips, in their view, is easing.
But we think a cool head is warranted. Running a demo on Horizon's toolchain and entering automotive-grade mass production are two different things. Automotive grade demands functional safety (ISO 26262, the automotive electronics safety certification standard), chip-level reliability across the full lifecycle, and deep coupling with the vehicle system — getting one model running proves none of these. There is another layer of risk: in on-device AI toolchains, Horizon must coexist long-term with the mature ecosystem of NVIDIA's Orin series. The contest is not just toolchain speed but also automakers' engineering habits and mass-production experience.
Impact on regular people
For enterprise IT and embedded teams: if your company works on autonomous driving, robotics, or industrial vision projects, domestic chips are no longer a backup option — they can now enter your formal cost and technical evaluation list.
For individual careers: on-device AI deployment engineers (covering model quantization and chip adaptation) will become a new talent gap in embedded AI, but the bar is high — you need to know both algorithms and hardware.
For consumer markets: in the short term you will not notice a difference in the car you buy; in the medium to long term, if the share of domestic chips in vehicles rises, vehicle BOM cost may fall, accelerating the price-down trajectory for smart cars in the RMB 100,000 tier.