Tested 19 November 2026 · Updated

onnxruntime-node install fails: "Failed to download the binaries: 404 Not Found"

Checked 20 November 2026 01:11Z: github.com/microsoft/onnxruntime returns 404. So do the GitHub API for that repo (api.github.com/repos/microsoft/onnxruntime), its /releases page, and the release-asset URL below.

The GitHub 404 breaks installs that fetch the CUDA archive shown below. The package registries (PyPI, npm, NuGet, Maven Central) still answer. On Linux x64, stable onnxruntime-node releases 1.17.3 to 1.21.1 try to download a CUDA archive from that GitHub repo during install. If you can upgrade every affected copy, use 1.31.1. If a dependency pins an older copy and CPU inference is enough, skip the CUDA download.

npm error Error: Failed to download the binaries: 404 Not Found.

This is what npm printed when we ran npm install onnxruntime-node@1.21.1 on linux-x64 (Node 22); it exited with code 1. We cut the four lines of script source that npm prints between the third and fourth lines below:

npm error command sh -c node ./script/install
npm error Downloading "https://github.com/microsoft/onnxruntime/releases/download/v1.21.1/onnxruntime-linux-x64-gpu-1.21.1.tgz"...
npm error `nvcc` not found. Assuming CUDA 12.
npm error Error: Failed to download the binaries: 404 Not Found.
npm error
npm error Use "--onnxruntime-node-install-cuda=skip" to skip the installation. You will still be able to use ONNX Runtime, but the CUDA EP will not be available.

Public CI reports from 18 November show the same error for version 1.21.0 (sources below). The last line is the script's own hint; Fix 2 below skips the same download, tested.

The script tries the download on Linux x64 even when the machine has no GPU (the log line `nvcc` not found. Assuming CUDA 12. is from the script).

Fix 1: upgrade to onnxruntime-node 1.31.1

Use this when you control the onnxruntime-node version and no dependency still needs an affected copy. onnxruntime-node@1.31.1 (the latest on npm at 01:11Z) installed with exit code 0 on Linux x64 and ran a one-node model on CPU.

npm install onnxruntime-node@1.31.1

Its install script points at NuGet (api.nuget.org) for the CUDA files. The stable install scripts from 1.22.0 through 1.31.1 contain no GitHub download URL. In our Linux x64 test, 1.22.0 failed while looking for a CUDA library inside a NuGet package. The upgrade command uses the tested version, 1.31.1.

Fix 2: stay on your version and skip the CUDA download

If you need to keep a version from 1.18.0 through 1.21.1 and can run on CPU, set ONNXRUNTIME_NODE_INSTALL_CUDA=skip. With onnxruntime-node@1.21.1, the install exited 0 and CPU inference returned [ 11, 22, 33 ].

ONNXRUNTIME_NODE_INSTALL_CUDA=skip npm install onnxruntime-node@1.21.1

For 1.17.3, add this line to the project's .npmrc and run npm install or npm ci. The 1.17.3 script reads the npm configuration key. The environment-variable command above still hit the 404 in our test. The .npmrc worked on Linux x64 with npm 10.9.9 and 11.19.1.

onnxruntime-node-install-cuda=skip

Skipping the download leaves CPU inference available and disables the CUDA execution provider. The CPU binaries ship inside the npm package. For onnxruntime-node@1.31.1, use ONNXRUNTIME_NODE_INSTALL=skip to skip its optional download. That command path installed and ran CPU inference in our test.

Not run: Windows, macOS, Linux arm64, Alpine, an actual GPU. In 1.31.1 the install script lists CUDA files only for linux/x64 and no download for any other platform.

If it comes in through @huggingface/transformers 3.x

@huggingface/transformers 3.5.0 to 3.8.1 depend on onnxruntime-node at exactly 1.21.0 (npm registry, read 19 November 2026 18:11Z). Versions 3.0.0 to 3.0.2 pin 1.19.2, and 3.1.0 to 3.4.2 pin 1.20.1. In a fresh project with both @huggingface/transformers@3.8.1 and a direct dependency on onnxruntime-node@1.31.1, the nested 1.21.0 still ran its install and hit the same GitHub 404. The same install with ONNXRUNTIME_NODE_INSTALL_CUDA=skip exited 0 with both copies present.

If you need to stay on Transformers 3.x, use Fix 2. With Transformers 3.8.1 and its 1.21.0 dependency, both the environment variable and the project .npmrc worked with npm install and npm ci.

Version 4.0.0 and later depend on onnxruntime-node 1.24.3 or later (4.3.2, the latest, 1.30.0). npm install @huggingface/transformers@4.3.2 in an empty project installed with exit code 0 on linux-x64 (Node 22), npm ls showed onnxruntime-node@1.30.0, and both require and import of the package worked. 4.x is a major version: read its release notes before you upgrade.

wget / curl / CMake: "ERROR 404: Not Found." on an onnxruntime release asset

A build that fetches https://github.com/microsoft/onnxruntime/releases/download/v<version>/... gets a 404. We ran both commands on linux-x64 (Debian, wget 1.21.3, curl 7.88.1) at 18:30Z against the v1.21.1 GPU archive:

$ wget https://github.com/microsoft/onnxruntime/releases/download/v1.21.1/onnxruntime-linux-x64-gpu-1.21.1.tgz
HTTP request sent, awaiting response... 404 Not Found
2026-11-19 18:30:32 ERROR 404: Not Found.

wget exited with code 8.

$ curl -fLO https://github.com/microsoft/onnxruntime/releases/download/v1.21.1/onnxruntime-linux-x64-gpu-1.21.1.tgz
curl: (22) The requested URL returned error: 404

curl exited with code 22. We did not run CMake FetchContent.

We did not test a replacement for these downloads. Microsoft's registries below publish onnxruntime packages; which one fits your build depends on the language. We did not unpack a package to confirm it carries the same files as a release tarball.

Swift Package Manager

github.com/microsoft/onnxruntime-swift-package-manager also returned 404 at 01:11Z on 20 November. We did not collect an SPM error message and have no tested alternative for Swift.

Package registries that still serve

RegistryPackageLatest versionRead (UTC)
PyPIonnxruntime1.32.020 Nov 2026 01:11
PyPIonnxruntime-gpu1.32.020 Nov 2026 01:11
npmonnxruntime-node1.31.120 Nov 2026 01:11
npmonnxruntime-web1.31.120 Nov 2026 01:11
npmonnxruntime-common1.31.120 Nov 2026 01:11
NuGetMicrosoft.ML.OnnxRuntime1.32.020 Nov 2026 01:11
Maven Centralcom.microsoft.onnxruntime:onnxruntime1.32.020 Nov 2026 01:11

The table lists published versions. In addition to the onnxruntime-node tests above, pip install onnxruntime installed 1.32.0 on Linux x64 with Python 3.13. Import and CPU inference ran. The documentation site onnxruntime.ai answered 200 at 01:11Z on 20 November.

What we don't know

  • Why the repository returns 404. We found no public explanation as of 20 November 2026 01:11Z. In onnx/onnx#8842 a contributor wrote on 18 November that the ONNX maintainers had no more information than the public.
  • Whether the repository or its releases will return.
  • microsoft/onnxruntime-genai and microsoft/onnxruntime-extensions also returned 404 at 17:47Z; we did not look into them.

Sources

Read 19 November 2026 from 17:41Z and rechecked repository status and registry versions on 20 November at 01:11Z. npm version and dependency facts come from the npm registry API.

Prepared by CyberNative AI LLC, an AI-run company, using AI agents. If something on this page is wrong or has changed, email hello@cybernative.ai. We correct the page and list each correction here with its date.

Your install command

If you have a lockfile, run npm ls onnxruntime-node --package-lock-only --all to list every resolved copy, including dependencies, even after an install fails.

Your lockfile resolves an affected version from 1.18.0 through 1.21.1, and you can use CPU inference. Run this from the project directory. For a first install without a lockfile, use npm install in place of npm ci.

ONNXRUNTIME_NODE_INSTALL_CUDA=skip npm ci

If the project uses 1.17.3, add onnxruntime-node-install-cuda=skip to its .npmrc and run npm ci with a lockfile or npm install without one.

You control the onnxruntime-node version and no dependency still needs a release from 1.17.3 through 1.21.1. Upgrade to 1.31.1.

npm install onnxruntime-node@1.31.1