Edge AI

Hi ST Support Team,We encountered a reproducible HardFault when running an object detection model on the STM32N6570-DK, and have completed extensive isolation testing. We believe this points to an issue in the STEdgeAI 4.0.1 / Neural-ART toolchain combined with the Getting Started reference application (concurrent camera/display/NPU usage), rather than a configuration mistake on our side.[Environment]- Board: STM32N6570-DK (with IMX335 camera)- Project: Getting Started Object Detection App (stm32ai-modelzoo-services)- Model: st_yoloxn (1 class, 416x416), Quantized (input: uint8 / output: int8)- Toolchain: STEdgeAI Developer Cloud v4.0.1 (forced by cloud; v4.0.0 is no longer offered, though the project's own README states the shipped model files were generated with v4.0.0)[Symptom]Build, flash, and boot-from-flash all complete without error, but the LCD stays black. Debugging via STM32CubeIDE shows a HardFault during NPU inference:HardFault_Handler()__LL_ATON_RT_IrqErr(irqs)   // ll_ato

Hi,I recently erased the external flash of my STM32N6570-DK board while experimenting with STM32CubeProgrammer, and I'd like to restore the original out-of-the-box demo (the multi-application menu demo with image classification / pose estimation / face detection etc.).I found the "n6-ai-demos" entry under Tools & Software on the product page, but clicking "Get latest" leads to a 404 page not found error.Could someone please:Confirm whether this link is currently broken on ST's side, or Share a direct working link / attach the "STM32N6570-DK Out-of-the-box demo binary" fileBoard info:Board: STM32N6570-DK, Revision ID Rev Z Device ID: 0x486Thanks in advance!  

LL_ATON_LIB_Concat falls back to 4× memcpy when the concat axis is not leftmost — 32% of a 4 ms real-time budget on STM32N6Hardware: NUCLEO-N657X0-Q (STM32N657X0H3Q), Cortex-M55 @ 800 MHz, Neural-ART. Tools: ST Edge AI Core 4.0, STM32Cube_FW_N6 V1.4.0, STM32CubeIDE 2.1.1. Application: real-time audio, hop = 192 samples @ 48 kHz = 4 ms = 3.2 M cycles, hard deadline, underrun must stay 0.1. Concat falls back to a byte loop, and it is one third of the budgetOur graph is a streaming TCN (int8, L=12, C=88, T=48) run per audio hop with two voices, so the tensor layout has the voice axis leftmost: [C][V][T]. The graph contains 12 Concat blocks (channel-wise taps).Epoch profiling (LL_ATON_EB_DBG_INFO, DWT cycle counter on the M55) showed those 12 blocks costing 1 019 501 cycles per inference — 32% of our entire 4 ms budget. No compiler option changed the number.The cause is in the runtime source: LL_ATON_LIB_Concat takes its fast DMA path only when the concatenation axis is the leftmost signif

Hello, I recently encountered this error while deploying an AI model on an STM32N647 for performance verification. I noticed that others have encountered the same issue on the official website, but following the suggested solutions did not resolve my problem—I’m still getting the same error. This issue has been troubling me for a week now. Are there any other ways to resolve it? 

Dear ST Edge AI Support Team,We are deploying a YOLO11n person-detection model on the STM32N6 NPU using the relocatable(LRUN) generation flow. While decoding the generated runtime buffer descriptors on-target, wefound that the top-level *input* user-IO buffer's channel-position metadata disagrees with itsown reported memory shape, and does not change when explicitly requesting a channel-last inputlayout.Configuration Target: STM32N6 ST Edge AI Core v4.0.1-20581 7ed50de05 Model: yolo11n_256_quant_pc_uf_od_coco-person-st.tflite (quantized, uint8 input, 256x256x3) stedgeai cli command (as echoed by network_generate_report.txt):   "stedgeai generate --model yolo11n_256_quant_pc_uf_od_coco-person-st.tflite --relocatable    split --target stm32n6 --no-inputs-allocation --no-outputs-allocation --optimization    balanced --st-neural-art user_profile@profile-neural-art.json --name network --verbosity 1    --c-api st-ai --inputs-ch-position chlast --workspace ./workspace --output ./outputs"Issue

I am integrating X-CUBE-AI and the STM32H7 Safety STL into the same STM32H743 Keil/ArmClang project.The X-CUBE-AI Keil runtime requires short enums and 16-bit wchar_t (-fshort-enums -fshort-wchar). However, the Safety STL library appears to require full-size enums and 32-bit wchar_t.Configuring the project for X-CUBE-AI causes the Safety STL library to fail linking. Configuring it for the Safety STL causes Arm linker ABI errors when linking the X-CUBE-AI runtime, including enum-size and wchar_t mismatches.Is there an officially supported compiler configuration or compatible library variant that allows X-CUBE-AI and the STM32H7 Safety STL to coexist in the same ArmClang application?

Hello,I’m using STM32Cube AI Studio to optimize and validate an ONNX model on a NUCLEO-N657X0-Q board. In general, everything works fine. However, I would like the board to be configured in overdrive mode (npu_freq = 1000MHz, mcu_freq = 800MHz) by Cube AI Studio for the validation runs on the target because I’m hoping to achieve a better inference time.When executing new “runs” in Cube AI Studio, the device attributes are always as seen in the screenshot below.Results overview of a run in Cube AI StudioIf I use the the ST developer cloud to optimize and benchmark my model with the STM32N6570-DK board (because the NUCLEO-N657X0-Q is not available in the DC), the device attributes are as follows:Device attributes for the benchmark in the developer cloudSince those two boards have the same STM32N657X0H3Q MCU, I expect a similar inference time on my NUCLEO-N657X0-Q board with Cube AI Studio if I could just activate the overdrive mode. I didn’t find any settings in the Cube AI Studio user i

Somesays ago, I successfully used STM32Cube.AI Studio to deploy a model to an STM32U585CIU6 board and test it.However, yesterday evening I upgraded STM32CubeIDE, STM32CubeProgrammer, and STM32CubeMX. After that, when I tried to run the same process again, I started getting an error.I searched online and found some similar cases. The suggested solution was to update all related software, so I updated the tools accordingly. However, the same error still occurs.I also tried using another laptop. On that laptop, I performed the same software upgrades and then tested with an STM32H743VIT6 board, but the same error occurred there as well.Because of this, I suspect that I may have missed updating some required software or package.Has anyone experienced this issue before or found a solution? I would greatly appreciate any advice or guidance.ps : stedgeai version is v4.0.1.   

我在尝试用STM32Cube AI Studio运行一个STM32F407ZGT6项目时遇到了“启动失败运行”错误，我不确定问题的原因。一开始运行正常，但之后错误一直出现。使用“主机”模式时，它正常工作。2026-07-13 10：49：46 错误：以下路径中找不到4.0.0版本的可执行文件：C：/*ST*/*STEdgeAI*/4.0，C：/*ST*/*STEdgeAI*/*2026-07-13 10： 49：47 错误：以下路径中找不到4.0.0版本的可执行文件：C：/*ST*/*STEdgeAI*/4.0，C：/*ST*/*STEdgeAI*/*2026-07-13 11：14：59 错误：从公共IP获取国家时出现错误：获取失败2026-07-13 11：15：00 错误：以下路径中找不到4.0.0版本的可执行文件：C：/*ST*/*STEdgeAI*/4.0，C：/*ST*/*STEdgeAI*/*2026-07-13 11：15：01 错误：以下路径中找不到4.0.0版本的可执行文件：C：/*ST*/*STEdgeAI*/4.0，C：/*ST*/*STEdgeAI*/*

In file included from Middlewares/ST/AI/Misc/Src/ai_device_adaptor.c:19:Middlewares/ST/AI/Misc/Inc/ai_device_adaptor.h:655:8: error: unknown type name 'UART_HandleTypeDef'; did you mean 'DMA_HandleTypeDef'?  655 | extern UART_HandleTypeDef UartHandle;      |        ^~~~~~~~~~~~~~~~~~      |        DMA_HandleTypeDefMiddlewares/ST/AI/Misc/Inc/ai_device_adaptor.h: In function 'port_io_get':Middlewares/ST/AI/Misc/Inc/ai_device_adaptor.h:664:12: error: implicit declaration of function 'HAL_UART_Receive' [-Wimplicit-function-declaration]  664 |   status = HAL_UART_Receive(&UartHandle, (uint8_t *)c, 1, timeout);      |            ^~~~~~~~~~~~~~~~Middlewares/ST/AI/Misc/Inc/ai_device_adaptor.h: In function 'port_io_write':Middlewares/ST/AI/Misc/Inc/ai_device_adaptor.h:676:12: error: implicit declaration of function 'HAL_UART_Transmit' [-Wimplicit-function-declaration]  676 |   status = HAL_UART_Transmit(&UartHandle, buff, count, HAL_MAX_DELAY);      |            ^~~~~~~~~~~~~~~~~make: *

Hi ST Community,I am working with an STM32N657 and would like to use the integrated NPU directly from my own C/C++ application.I understand that X-CUBE-AI together with the Neural-ART can generate the code required to execute a neural-network model on the STM32N6 NPU.However, I would like to understand whether it is possible to use the NPU without X-CUBE-AI.Does STM32CubeIde provide HAL APIs that allow an application to directly initialize, configure, load, and start the STM32N6 NPU?

 Hello everyone,I am encountering errors when running Semantic Segmentation, Object Detection, and Pose Estimation models on my STM32N6570-DK board using STM32 Model Zoo Services. Here are the details of my setup and the errors I am facing:Hardware and Versions:Board: STM32N6570-DKModels: Semantic Segmentation, Object Detection, Pose EstimationTools: STM32Cube.AI 4.0.0, GCC 14.3.1, HAL 1.3.0, STMCubeIDE 2.1.1I followed the README instructions from the STM32 Model Zoo Services repository to test the mentioned models. I compiled and flashed the firmware onto the STM32N6570-DK board.Errors Encountered: plaintextCopyBus interface interrupt BUSIF0 ERR: 0x0 BUSIF1 ERR: 0x20 Epoch Controller ERROR interrupt: EC_IRQ = 0x0000000a Epoch Controller opcode counter: 0x00000000 Epoch Controller label: 0x00000000 ATON_STD_IRQHandler()@871: irqs=0xlx assertion "0" failed: file "[...]/stm32ai-modelzoo-services/application_code/.../ll_aton_runtime.c", line 875, function: __LL_ATON_RT_IrqErr&nb

HelloST Edge AI: Guide for STM32MPU 、 Section 3.2.1. Referring to “Generate NBG from TensorFlow™ Lite or ONNX,”I converted the .tflite file downloaded from ModelZoo to .nb, but I’m getting an error.$ /opt/ST/STEdgeAI/4.0/Utilities/linux/stedgeai generate -m ./yolov8n_256_quant_pc_uf_od_coco-person-st.tflite --target stm32mp25ST Edge AI Core v4.0.1-20581 7ed50de05E102(CliArgumentError): stm32mp25 not recognized as target. Supported targets areispu --device DEVICEmlcstm32f4stm32l4stm32l4pstm32g4stm32f3stm32wbstm32n6 [--st-neural-art]stm32f0stm32l0stm32g0stm32c0stm32wlstm32v8stm32u3 [--hsp]stm32h7stm32f7stm32u5stm32l5stm32h5stm32c5stm32mp157cIf you change the target name slightly, you'll get the following error: $ /opt/ST/STEdgeAI/4.0/Utilities/linux/stedgeai generate -m ./yolov8n_256_quant_pc_uf_od_coco-person-st.tflite --target stm32mp2ST Edge AI Core v4.0.1-20581 7ed50de05E102(CliArgumentError): stm32mp2 is not availableCould you tell me how to convert a .tflite file to

hey i am struggling to flash by custom onnx model on stm32N6570-DK . I tired using the stm32 cube ai cloud website but it only gives me option to generate c code and not a project or elf file . So i shifted to stm32cube Ai studio but it keeps me throwing this error -  (*) 'input'/'output' buffers are allocated in the activations bufferRunning the ONNX model...Running the ST.AI c-model (AI RUNNER)...(name=network, mode=TARGET)E200(ValidationError): TARGET: Unable to bind the ST.AI runtime with "network" c-model: [] E801(HwIOError): Invalid firmware - COM3:115200 can anyone help me to get started with the same

Hi ST Community, I am trying to deploy a YOLO-FastestV2 INT8 object detection model on the STM32N6570-DK. The model is successfully converted and runs on the board, but I am facing issues with the detection results.Model InformationModel: YOLO-FastestV2 (INT8) Number of classes: 6 Parameters: 235,120Model InputInput_0_out_0uint8 [1,1,320,320]Model OutputsQuantize_297_out_0int8 [1,12,20,20]Quantize_313_out_0int8 [1,3,20,20]Quantize_311_out_0int8 [1,6,20,20]Quantize_330_out_0int8 [1,12,10,10]Quantize_346_out_0int8 [1,3,10,10]Quantize_344_out_0int8 [1,6,10,10]Issues I am FacingThe bounding boxes are not displayed at the correct positions on the camera image. The detected bounding boxes appear to be shifted or incorrectly scaled. Even when the camera is not viewing any object (or the camera feed is unclear), random bounding boxes continue to be displayed. These false detection remain on the screen even though there should be no valid detection.Verification PerformedI also tested the quanti

Hi,I am working on a project using STM32H743ZITx and trying to deploy an ONNX model using X-CUBE-AI.My .ld file was doing just fine until I add the X-CUBE-AI in the middleware. It will split the BSS range across DTCMRAM and AXI‑RAM. The startup file’s BSS‑zeroing logic accesses invalid non‑hardware‑backed addresses between the two RAM regions and results in a HardFault right after Reset_Handler.My configuration is:MCU: STM32H743ZITx STM32CubeMX: 6.18.0 STM32Cube FW_H7: 1.13.0 STM32CubeIDE for Visual Studio Code: 3.9.0 Enabled X‑CUBE‑AI 10.2.1Files before and after the configuration of X-CUBE-AI are attached.

Hi all,i have a small question about Validation of a trained model.Atm i only work on PC with NEAI Studio Version 5.1.1.I used only data out of a *.csv fileI trained (benchmark) the model with around 170 Lines and 16 axis to get an accuracy of ~97%.I then validated the library with some test data with same axis count and data preparation. The learning file has the minimum amount of lines as suggested from the benchmark, i tried more lines with same result.The validation shows that even good data is treated as abnomaly so i only get an accuracy of max. 50%.Are there any hints what i am doing wrong? Many thanks in advance for answers.BR’sMartin

I am trying to generate a program a run using STM32Cube AI Studio and the STM32N6570-Dev Kit, but am running into an issue with the run failing due to a resource busy or locked (the .cproject).  Some steps I’ve already taken:Closed all other STM programs (IDE, Programmer, MX) Completely deleted and made a new project Confirmed all ST programs (IDE, Programmer, MX) are up to date by cleaning and re-installing from ST website Confirmed all ST programs are linked to the AI Studio and confirmed they’re all downloaded in their default location. Restarted my PC.Notes:I am on Windows 11, Version 10.0.26200 Build 26200 Model = ST YOLO-X Nano - Person Detection. Source: Pre-trained model from ST Model Zoo Path: stm32ai-modelzoo/object_detection/st_yoloxn/ST_pretrainedmodel_public_dataset/coco_2017_person/st_yoloxn_d033_w025_256/

STM32N657 / STEdgeAI 4.0.1 - Non-deterministic inference results with a minimal STAI wrapperHello,I am currently evaluating STEdgeAI 4.0.1 on an STM32N657 before integrating AI support into my RTOS (uKOS-X).My objective is not to use the generated CubeMX application. Instead, I want to build a small self-contained library that encapsulates all ST Edge AI resources and exposes only a minimal API to the application.The wrapper API is intentionally very small:gan_npu_status_t npu_init(void);gan_npu_status_t npu_putInput(const void *input);gan_npu_status_t npu_inference(void);gan_npu_status_t npu_getOutput(void *output);The application never accesses any STAI object directly.This is only a first validation step, therefore I intentionally use synchronous execution (STAI_MODE_SYNC) and polling only. Interrupt-driven execution will be implemented later.Configuration STM32N657 STEdgeAI 4.0.1 Neural-ART generated network Single network instance ST-generated sources left completely uncha

I would like to clarify the data constraints for Benchmark workflow in NanoEdge AI Studio:What is the maximum supported number of lines for each imported CSV dataset during Benchmark?Is there any official documented hard limit on the number of rows, or only practical recommendations?Some community references mention a rough threshold of ~100,000 lines, and files exceeding this size may cause memory overflow, import truncation or failed Benchmark launch. Could you confirm whether this threshold is formally defined?

Hello ST Community, I noticed that when using the benchmarking tool, datasets larger than 250 MB are automatically reduced using a "representative sampling strategy" that preserves the diversity and distribution of the original data. I would like to understand more details about this process: 1. What specific sampling algorithm/strategy is used? (e.g., random sampling,    stratified sampling, systematic sampling, etc.) 2. How is the target size of 250 MB determined? Is this configurable? 3. How does the tool ensure that the sampled dataset preserves:   - The original data distribution?   - The diversity of data samples?   - Edge cases and boundary values? 4. After optimization, when we validate the model on the full dataset, what    performance differences should we typically expect between the sampled    dataset and the full dataset? 5. Is there any documentation that describes this sampling mechanism in detail? Any information or pointers to relevant documentation would be greatly ap

how can we go initial stm32N6?

hi all, I am trying to flash + run a relocatable / runtime loadable model. The problem is that the inference results are wrong (they never change, no matter what the buffer contains when I start the inference). If I use the stai_model files from the AI Cube Studio, the model works correctly, but is not relocatable… Am aware of matching the address where I flash the network_rel.bin and generating the model. (← see URL) I am using these docs. (← see URL) I am enabling all AXISRAM3...6 I generate the model as follows:stedgeai.exe generate -m "path/to/model.tflite" --input-data-type uint8 --no-inputs-allocation --no-outputs-allocation --target stm32n6 --st-neural-art --reloc# final output something like this:+--------------------------+-------------------------------+------+------------+--------+| name (addr) | flags | foff | dst | size |+--------------------------+-------------------------------+------+------------+--------+| xSPI2 (0x20009332

I am using seeed studio AI assistant board and I am not able generate the cube  project using StudioI am facing Memory Overflow Issue Is there any way to solve thisAnd Seeed Board is not supported when Board is selected during project creation 

I’m trying to convert an ONNX model which has a Conv layer that takes weights as input (has 2 inputs). Edge AI gives me:NOT IMPLEMENTED: Convolution with multiple not constant inputs not supportedAre there any short-term plans to support such layers in ST Edge AI?I also tried to change the name of the layer to some custom name (DynConv). The error in ST Edge AI changed to:NOT IMPLEMENTED: Unsupported layer types: DynConvI thought that I could bypass this error by using --custom with a JSON definition, but the JSON format is poorly documented. Also, many attempts at different formats seem to indicate that --custom only works for Keras. Is this correct?