# ONNX Runtime

11 Hacker News commenters recommended ONNX Runtime and 13 criticized it, November 2025 to November 2026. It was recommended in 8 of 13 months.

**Verdict:** Recommended as a fast, well-optimized engine for running models on CPU and in browsers. Main caveat: converting TensorFlow and PyTorch models often fails.

## Why people recommend it

- Well optimized for CPU inference (2 comments): "ONNX is my first suggestion to people looking for speed gains on CPU" ([source](https://news.ycombinator.com/item?id=48771356)) "ONNX is incredibly well optimized for CPU, best in class even" ([source](https://news.ycombinator.com/item?id=48965716))
- Runs in browsers and production (2 comments): "Impressive work on making this run entirely in-browser with ONNX Runtime Web!" ([source](https://news.ycombinator.com/item?id=46840210)) "We currently serve 5+ million inference requests per day" ([source](https://news.ycombinator.com/item?id=48459976))
- Engine behind other tools (3 comments): "everything is enabled by onnx with some cgo (or Rust) plumbing on top" ([source](https://news.ycombinator.com/item?id=48458538)) "I know they have a solid ONNX engine" ([source](https://news.ycombinator.com/item?id=46858019)) "ported to plain C from ONNX Runtime" ([source](https://news.ycombinator.com/item?id=49151149))

## What people warn about

- Large size and heavy dependencies (4 comments): "The ONNX Runtimes are also huge (150mb), JNI, and drag gargantuan dependencies!" ([source](https://news.ycombinator.com/item?id=47745559)) "the primary onnxruntime jar a boat anchor; weighing is 90mb+ just by itself" ([source](https://news.ycombinator.com/item?id=47746013)) "In my case I found that ORT alone was 3.4MB over the wire" ([source](https://news.ycombinator.com/item?id=48971236))
- Model conversion often fails or is painful (3 comments): "Conversions from models written with Tensorflow and Pytorch often fail." ([source](https://news.ycombinator.com/item?id=46384579)) "Exporting models to work on ONNX is a pain in the ass." ([source](https://news.ycombinator.com/item?id=46552703)) "an incredible inconvenience masked as the easy way of doing things" ([source](https://news.ycombinator.com/item?id=48460228))
- Wrong outputs or slow quantization (3 comments): "That onnx model run using onnxruntime with cuda ep is a different model than the one run with TRT ep." ([source](https://news.ycombinator.com/item?id=48249562)) "never got it produce correct outputs with my quantized models" ([source](https://news.ycombinator.com/item?id=49746341)) "ORT quantizer made it slower than FP32" ([source](https://news.ycombinator.com/item?id=48313441))
- Crashes and opaque behavior (2 comments): "unable to fix it. latest dev ort, dispoing tensors, restarting the session/worker. nothing, crashes every time." ([source](https://news.ycombinator.com/item?id=49062527)) "They are incredible for production, but they are "Blackboxes."" ([source](https://news.ycombinator.com/item?id=46812588))

## Alternatives hackers recommend

- [MTPLX](https://hackerapproved.com/p/mtplx.md): 15 recommend
- [Tiny-vLLM](https://hackerapproved.com/p/tiny-vllm.md): 11 recommend
- [OpenVINO](https://hackerapproved.com/p/openvino.md): 10 recommend
- [ONNX](https://hackerapproved.com/p/onnx.md): 9 recommend
- [Timber](https://hackerapproved.com/p/timber.md): 9 recommend

## Cite as

Hacker Approved, "ONNX Runtime", https://hackerapproved.com/p/onnx-runtime/ (counts from 1,795,640 Hacker News comments, November 2025 to November 2026).