New Release: Ultralytics v8.4.120

Discussion
, ,
1

:rocket: Ultralytics v8.4.120 Released

Ultralytics v8.4.120 delivers cleaner deterministic CUDA training, more reliable TensorFlow export setup, and expanded documentation for LLM workflows and AI coding-agent integrations. This release focuses on robustness and developer experience, with no major model architecture changes.

:glowing_star: New Features

:robot: OpenAI-Compatible LLM Interface

The new LLM documentation covers a consistent OpenAI-compatible interface for:

  • Text and image inputs
  • Streaming and asynchronous workflows
  • Provider-specific configurations
  • Combined Ultralytics YOLO and LLM pipelines
  • OpenAI-compatible services including DeepSeek, Kimi, Z.AI GLM, OpenRouter, and local servers

The documented and runtime default model is now gpt-5.6-luna.

This work was contributed by @onuralpszr in PR #25789.

:puzzle_piece: Agent Skills Integration Guide

New documentation introduces the official Ultralytics Agent Skills repository, providing structured guidance for AI-assisted workflows such as:

  • Model selection
  • Dataset preparation
  • Training and tuning
  • Inference
  • Model export

The guide also includes installation instructions for Claude Code, Codex, and other compatible coding agents.

This first-time contribution was made by @JaviChulvi in PR #25787. Welcome to the community! :tada:

:high_voltage: Improvements

Deterministic CUDA Anchor Generation

Detection anchor creation now uses deterministic arange-based generation instead of CUDA cumulative-sum operations.

This change:

  • Removes recurring cumsum_cuda_kernel warnings during deterministic training
  • Keeps training logs cleaner and easier to inspect
  • Preserves runtime device handling for traced and TorchScript GPU models
  • Avoids incorrectly fixing device information during tracing

The improvement was implemented by @glenn-jocher in PR #25806.

:hammer_and_wrench: More Reliable TensorFlow Exports

TensorFlow and non-YOLO export dependency installation no longer relies on the obsolete NVIDIA package index. Since onnx-graphsurgeon is available directly from PyPI, this should reduce DNS and connectivity failures in:

  • CPU-only environments
  • CI pipelines
  • Restricted networks
  • Isolated export setups

:package: Upgrade Now

Install or upgrade to the latest release with:

pip install -U ultralytics

Confirm the installed version with:

yolo checks

Review the complete package on the Ultralytics v8.4.120 release page, or explore every change in the v8.4.119 to v8.4.120 full changelog.

Give v8.4.120 a try and share your feedback, export results, or any issues you encounter. Thanks to the Ultralytics team and YOLO community for continuing to improve the ecosystem! :raising_hands: