A YOLO that benchmarks itself on T4s and mid-range phones

What it does

YOLOv6 is a single-stage object detector built by Meituan for industrial deployment. It offers a tiered family of models—from the compact YOLOv6-N up to the high-resolution YOLOv6-L6—trained with self-distillation and benchmarked on COCO val2017. The repository also hosts branches for segmentation, face detection, and a YOLOv6Lite sub-family aimed at mobile CPUs.

The interesting bit

While most detection repos lead with mAP, this one leads with a full deployment matrix: TensorRT fp16 throughput on T4 GPUs, INT8 quantized accuracy drops, and MNN inference latency across Qualcomm, Dimensity, and older Snapdragon chips. The implicit message is that the model was trained to be deployed, not just cited.

Key highlights

• Server benchmarks cover batch-1 and batch-32 TensorRT fp16 speeds on a T4, alongside params and FLOPs for every variant.
• Quantized RepOpt checkpoints are published, so you can see the exact INT8 speedup and mAP cost before committing to quantization.
• YOLOv6Lite scales down to 224×128 input and publishes MNN timings on high-, mid-, and low-end mobile chips.
• Most checkpoints are trained with self-distillation; the exceptions (YOLOv6-N6 and YOLOv6-S6 at 300 epochs) are explicitly noted.
• Satellite branches add segmentation and face-detection heads without abandoning the core training pipeline.

Caveats

• The README is rich with tables but thin on architecture; you will need the arXiv papers for design rationale.
• Mobile numbers are specific to MNN 2.3.0 with arm82 acceleration, so expect different latencies on other runtimes.

Verdict

Worth a look if you need detection weights that come with pre-measured latency budgets for both server GPUs and cheap phones. If you are after novel architectural research, the papers are the better starting point than the repo.