Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models

Abstract: Real-time vision demands models that are accurate, efficient, and simple to deploy across diverse hardware. The YOLO family has become widely deployed for this reason, yet most YOLO detectors still rely on non-maximum suppression at inference, carry heavy detection heads due to Distribution Focal Loss, require long training schedules, and can leave the smallest objects without positive label assignments. We present Ultralytics YOLO26, a unified real-time vision model family that addresses these limitations through coordinated architecture and training advances. YOLO26 uses a dual-head design for native NMS-free end-to-end inference and removes DFL entirely, yielding a lighter head with unconstrained regression range. Its training pipeline combines MuSGD, a hybrid Muon-SGD optimizer adapted from LLM training; Progressive Loss, which shifts supervision toward the inference-time head; and STAL, a label assignment strategy that guarantees positive coverage for small objects. Beyond detection, YOLO26 introduces task-specific head and loss designs for instance segmentation, pose estimation, and oriented detection, producing consistent gains across tasks and scales. The family spans five scales (n/s/m/l/x) and supports detection, instance segmentation, pose estimation, classification, and oriented detection in a single pipeline, with an open-vocabulary extension, YOLOE-26, for text-, visual-, and prompt-free inference. Across all scales, YOLO26 achieves 40.9-57.5 mAP on COCO at 1.7-11.8 ms T4 TensorRT latency, advancing the accuracy-latency Pareto front over prior real-time detectors, while YOLOE-26x reaches 40.6 AP on LVIS minival under text prompting. Code and models are available at https://github.com/ultralytics/ultralytics.

• The paper presents a novel dual-head, NMS-free model that replaces DFL with unconstrained regression, improving large-object localization and reducing computational overhead.
• It employs Progressive Loss and the MuSGD optimizer to balance one-to-many and one-to-one training objectives, resulting in higher mAP and faster convergence.
• The architecture supports extensive task extensions—instance segmentation, pose estimation, and open-vocabulary detection—with robust deployment across 19 export backends.

Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models

Architectural Advancements in YOLO26

YOLO26 represents a comprehensive overhaul of YOLO-family detectors, targeting sustained improvements in accuracy, efficiency, and deployment generality. Central to this iteration is the migration to a dual-head, NMS-free inference design, which enables both direct, fixed-cardinality end-to-end (E2E) decoding and a fallback dense branch with conventional NMS. This rearchitecting is supported by three principal innovations: the full removal of Distribution Focal Loss (DFL) in favor of unconstrained regression, the introduction of a Muon-based gradient optimizer (MuSGD), and two training mechanisms—Progressive Loss and Small-Target-Aware Label Assignment (STAL)—to optimize supervisory signal allocation and guarantee small object coverage, respectively.

The overall model schematic reflects these priorities, with the backbone and neck shared across tasks feeding into the dual detection heads. (Figure 1)

Figure 1: Training pipeline of Ultralytics YOLO26. The shared backbone and neck feed the dual detection heads, STAL improves assignment robustness for tiny targets, Progressive Loss reweights the one-to-many and one-to-one objectives over training, and MuSGD performs the final parameter updates.

Further, the modular head design permits streamlined task extensions—instance segmentation via multi-scale prototype masking, pose estimation through uncertainty-aware keypoint regression, and oriented bounding box (OBB) detection using enhanced angle parameterization.

The deployment pipeline supports both E2E and NMS-based inference paths, with automatic fallback determined by target backend capabilities, maximizing deployment universality. (Figure 2)

Figure 2: Deployment pipeline of Ultralytics YOLO26. A trained model supports two inference paths: the default one-to-one NMS-free path and the optional one-to-many path with NMS, while preserving compatibility with a broad set of export targets.

Methodological Innovations: DFL Removal, Optimizer, and Supervision

DFL Removal & Direct Regression: YOLO26 eliminates DFL, reducing per-head parameters (e.g., 12% fewer parameters, 20% fewer FLOPs for the baseline n model). This circumvents DFL’s limited regression range, which constrains large-object localization at high resolutions and forces computational bloat with increased bin counts. Restoration of localization fidelity, especially at 1280 resolution, is demonstrated both numerically (AP improvement up to +1.3, AP$_L$ up to +2.2 without DFL) and qualitatively. (Figure 3)

Figure 3: Qualitative comparison for large-object localization at 1280 resolution. From left to right, each row shows predictions from the model with DFL, predictions from the counterpart without DFL, and the ground-truth annotations. The DFL-free head better preserves full-object extent on large targets, supporting the quantitative gains reported.

Progressive Loss: The model addresses the one-to-one/one-to-many branch optimization imbalance by dynamically annealing the loss weighting during training, smoothly shifting emphasis from dense (early-stage) to E2E (late-stage) supervision. Optimal scheduling initiates with a one-to-many/one-to-one mix of (0.8, 0.2) and transitions linearly to (0.1, 0.9), yielding measurable AP increases over static weighting.

MuSGD Optimizer: A hybrid optimization regime combines Muon’s learnable-orthogonalized updates with SGD-momentum, applying type-specific partitioning (pure SGD for 1D params, hybrid for high-dimensional weights). Empirically, MuSGD increases mAP by +0.4 with a 16.7% reduction in epochs compared to vanilla SGD, and consistently yields improvements in transfer to classification settings.

STAL: STAL resolves the well-known “TAL zero assignment” artifact, in which small GT boxes receive no positive anchors post-downsampling. By inflating candidate boxes for assignment purposes but preserving original regression targets, STAL guarantees non-trivial gradient signal for the smallest objects—objectively improving AP$_S$ and overall system sample efficiency. (Figure 4)

Figure 4: Qualitative comparison between TAL baseline (left), STAL with $s_{\mathrm{ref}=16$ (middle), and ground-truth annotations (right) on COCO validation images. STAL improves small-object detection by ensuring sufficient anchor coverage for tiny ground-truth boxes.

Task-Specific Extensions

The architecture seamlessly integrates task heads and loss functions for five core vision problems:

• Multi-task Backbone: A unified backbone/neck, detailed architecturally in (Figure 5), drives all downstream heads, with modular blocks (Figure 6) supporting both low-latency and high-capacity variants.
• Instance Segmentation: A multi-scale prototype module aggregates pyramid-level features for robust shared mask bases. Auxiliary dense semantics are fed only during training, benefiting mask AP without deployment impact.
• Pose Estimation: Residual Log-Likelihood Estimation replaces scale-invariant loss with a flow-based probabilistic model, introducing explicit uncertainty regression for each keypoint.
• OBB Detection: A long-edge, aspect-ratio-aware angle definition reduces edge-swapping discontinuities and is paired with a custom angle loss for accurate prediction of both highly elongated and near-square objects. Qualitative comparison unambiguously shows YOLO26's improved angle prediction on square instances relative to YOLO11. (Figure 7)
• Open-Vocabulary Detection/Segmentation (YOLOE-26): YOLO26 is the backbone for YOLOE-26, which supports text-prompted, visual-prompted, and prompt-free operations, leveraging an upgraded MobileCLIP2 encoder and pseudo-label data refinement.

Figure 7: Qualitative comparisons between YOLO26x-obb and YOLO11x-obb models for square rotated objects detection. Green boxes are ground truth, while red boxes indicate predictions. YOLO26 demonstrates better angle predictions over YOLO11 on square objects.

Figure 5: Ultralytics YOLO26 architecture diagram. The figure summarizes the end-to-end model structure, including the shared backbone and neck, the task heads, and the high-level information flow across the network.

Figure 6: Building blocks used to assemble the Ultralytics YOLO26 architecture. The figure details the constituent modules that compose the full network.

Numerical Results and Performance Analysis

YOLO26 sets new state-of-the-art accuracy–latency trade-offs across multiple tasks and model scales. On COCO detection, mAP ranges from 40.9 (n) to 57.5 (x) with T4 TensorRT latency from 1.7 ms to 11.8 ms. Direct comparisons with strong transformer-based real-time detectors (e.g., D-FINE, DEIM, RT-DETR) and recent YOLO baselines show consistent improvements at both the medium and large end of the Pareto frontier. E2E inference trails NMS-based prediction by <0.8 AP, offering clean deployment without significant accuracy penalty.

On task benchmarks, YOLO26 yields:

• Segmentation: +2.4–3.7 mask AP over YOLO11 (COCO)
• Pose Estimation: +2.1–7.2 AP (COCO keypoints)
• OBB Detection: +2.5–3.4 mAP (DOTA-v1.0); AP$_{75}$ gains of 4.6–6.0 highlight enhanced localization
• Open-Vocabulary (YOLOE-26): Achieves up to 40.6 AP on LVIS minival (text-prompted), exceeding DetCLIP-T by 6.2 AP

All models maintain full support for 19 export backends (ONNX, TensorRT, CoreML, NCNN, etc.), with up to 43% faster CPU inference and reduced model footprint relative to YOLO11.

Implications and Future Directions

YOLO26 fuses dual-head E2E deployment, operator-standard regressor designs, robust optimizer schedules, and architecture-task unification to produce a versatile foundation for both research and industry pipelines. Notably, it demonstrates that head parameter reduction and native NMS elimination do not necessitate AP sacrifice, provided training interventions are carefully designed.

By making core design choices (e.g., DFL removal, dedicated optimizer, label assignment for small objects) explicit and modular, YOLO26 is positioned for future extensibility—potential avenues include adaptive loss annealing, transfer learning from web-scale data, and architectural enhancements specialized for multimodal or grounding-intensive pretraining corpora.

YOLO26's open-vocabulary extension, YOLOE-26, further illustrates the adaptability of the architecture for emerging vision–language benchmarks without compromising real-time constraints. The data engine for pseudo-label augmentation and improvements in prompt alignment (MobileCLIP2) indicate that distributional supervision and textual grounding will be central to ongoing improvements.

Conclusion

YOLO26 operationalizes systematic architecture and training improvements—a DFL-free dual-head design, MuSGD, Progressive Loss, and STAL—across five scales, driving both theoretical and practical advances over antecedent YOLO models. Its unified, multitask-capable architecture and open deployment support underpin its suitability as a reference baseline for real-time vision going forward. The strong empirical results, versatility across tasks, and robust open-vocabulary capabilities set a high bar for subsequent real-time vision model development (2606.03748).