# New Release: Ultralytics v8.4.81

**URL:** <https://community.ultralytics.com/t/new-release-ultralytics-v8-4-81/2087>\
**Category:** Discussion\
**Tags:** releases, announcements, ultralytics-official\
**Created:** [June 28, 2026, 1:33pm UTC](https://community.ultralytics.com/t/new-release-ultralytics-v8-4-81/2087 "2026-06-28T13:33:04Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![glenn-jocher](https://sea1.discourse-cdn.com/flex001/user_avatar/community.ultralytics.com/glenn-jocher/32/83_2.png) [@glenn-jocher](https://community.ultralytics.com/u/glenn-jocher)\
**Post date:** [June 28, 2026, 1:33pm UTC](https://community.ultralytics.com/t/new-release-ultralytics-v8-4-81/2087/1 "2026-06-28T13:33:04Z")

</div>

# Ultralytics v8.4.81 is out 🚀

**Quick summary:** Ultralytics YOLO `v8.4.81` is a polish-and-reliability release focused on **FastSAM stability** , **safer and clearer export precision behavior** , better **YOLO26 export coverage** , and a handful of platform and integration fixes that make day-to-day workflows smoother.

If you use FastSAM, export models often, or are deploying YOLO26 across different runtimes, this is a great update to pick up. You can explore the release on the [v8.4.81 release page](https://github.com/ultralytics/ultralytics/releases/tag/v8.4.81) and review every change in the [full changelog](https://github.com/ultralytics/ultralytics/compare/v8.4.80...v8.4.81).

## 🌟 Highlights

### 🧠 FastSAM multiple text prompts no longer crash

A key fix in this release resolves a `ValueError` in `FastSAMPredictor.prompt()` when passing **multiple text prompts**.

This affected common prompt-based segmentation workflows, especially when smaller masks were filtered before prompt matching. With the new vectorized index remapping, prompt lists such as `["a photo of a dog", "a photo of a person"]` now work as expected.

Implemented in [PR #24952 by @ahmet-f-gumustas](https://github.com/ultralytics/ultralytics/pull/24952) from [@ahmet-f-gumustas](https://github.com/ahmet-f-gumustas) 🙌

### 📦 Export precision handling is now clearer and stricter

Export behavior has been cleaned up to use `quantize=` consistently, replacing older internal reliance on `half=True` and `int8=True` state.

That means unsupported precision requests now **fail clearly** instead of silently changing to another precision. A few examples:

- CoreML FP32 stays FP32 when requested
- CPU TorchScript FP16 now raises a clear error
- Format-specific precision support is better documented

This work landed across:

- [PR #24950 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24950) from [@glenn-jocher](https://github.com/glenn-jocher)
- [PR #24951 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24951) from [@glenn-jocher](https://github.com/glenn-jocher)

### 🤖 More YOLO26 export smoke-test coverage

YOLO26 is now used more broadly as the main export smoke-test model in CI for several formats, improving confidence around the latest recommended model family.

Included in [PR #24957 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24957) from [@glenn-jocher](https://github.com/glenn-jocher)

## Improvements

### 📘 Export docs got a major accuracy pass

Deployment docs were updated to better explain precision support across formats including CoreML, TorchScript, OpenVINO, MNN, RKNN, TensorFlow formats, ExecuTorch, and Sony IMX500.

This should reduce guesswork and make export decisions much easier for deployers.

Included in [PR #24950 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24950) from [@glenn-jocher](https://github.com/glenn-jocher)

### 🧬 Embedding behavior is more intuitive

`model.embed()` and `model.predict()` now play more nicely together. After embedding extraction, a normal prediction call returns standard `Results` again instead of unexpectedly remaining in embedding mode.

Included in [PR #24958 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24958) from [@glenn-jocher](https://github.com/glenn-jocher)

### ⚡ Safer imports for standard YOLO workflows

SAM/SAM3-related imports are now lazier, helping avoid import-time issues in environments that only need standard YOLO tasks and not SAM dependencies.

This was part of the general reliability work in [PR #24949 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24949) from [@glenn-jocher](https://github.com/glenn-jocher)

### 🪟 Better Windows support for ONNXRuntime C++

The ONNXRuntime C++ example now handles non-ASCII and UTF-8 Windows paths more reliably through safer filesystem path conversion.

Included in [PR #24953 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24953) from [@glenn-jocher](https://github.com/glenn-jocher)

### 🧩 Axelera export reliability improvements

Axelera-related export handling was tightened with:

- a `numpy<=2.3.5` pin to avoid broken YOLO26 attention-graph exports
- stronger export output validation so suspiciously tiny outputs fail loudly

Covered in:

- [PR #24947 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24947) from [@glenn-jocher](https://github.com/glenn-jocher)
- [PR #24956 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24956) from [@glenn-jocher](https://github.com/glenn-jocher)

## Documentation updates

### 📄 Hailo docs clarified

The Hailo docs now more clearly explain that HEF generation is currently a **manual ONNX-to-HEF workflow** , rather than a direct `model.export(format="hailo")` target.

The notes were also expanded around hardware and SDK compatibility, plus YOLO11 and YOLO26 guidance.

Updated in:

- [PR #24954 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24954) from [@glenn-jocher](https://github.com/glenn-jocher)
- [PR #24955 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24955) from [@glenn-jocher](https://github.com/glenn-jocher)

## Bug fixes and maintenance

- Sentry-driven reliability fixes in [PR #24949 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24949) from [@glenn-jocher](https://github.com/glenn-jocher)
- Reduced slow CI Jetson TensorRT coverage for faster, more efficient pipelines in [PR #24959 by @glenn-jocher](https://github.com/ultralytics/ultralytics/pull/24959) from [@glenn-jocher](https://github.com/glenn-jocher)

## Why this release matters 🎯

This update is all about making Ultralytics workflows feel more predictable and production-ready:

- ✅ **FastSAM prompting is more reliable**
- ✅ **Exports are more trustworthy and transparent**
- ✅ **YOLO26 gets stronger deployment coverage**
- ✅ **Embedding and prediction flows are less surprising**
- ✅ **Platform-specific edge cases are handled better**
- ✅ **Docs are clearer for real-world deployment decisions**

## Try it out

If you want to upgrade right away, you can install the latest package with:

```bash
pip install -U ultralytics

```

And if you are starting a new project, we recommend using [Ultralytics Platform](https://platform.ultralytics.com) together with [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) for the best current training and deployment experience.

## Share your feedback 💬

Please give `v8.4.81` a try and let us know how it works for your workflows, especially around **FastSAM prompting** , **export precision** , and **YOLO26 deployment**. Feedback, bug reports, and edge cases help the whole YOLO community improve faster.

Thanks to everyone contributing fixes, testing exports, and sharing reproducible reports ❤