fastembed-rs 7.4 makes ONNX Runtime optional, so Qwen3-only Candle builds stop linking ort

fastembed-rs, the Rust crate for local text and image embeddings and reranking, shipped five releases on 23 October. The one that changes how you build it is 7.4.0: ONNX Runtime is now optional.

The problem it fixes

Until now ort and ort-sys, the ONNX Runtime bindings, were unconditional dependencies. PR #335 shows what that meant for a crate that only wanted Qwen3 on the Candle backend, with default-features = false, features = ["qwen3", "metal"]: it still compiled ort, and its test binaries and executables failed to link unless you added ort-sys with disable-linking or pointed ORT_LIB_LOCATION at a runtime the code never called.

What changed in 7.4.0

• A new onnx feature, on by default. ort, ndarray, safetensors and the ONNX-based modules (TextEmbedding, SparseTextEmbedding, Bgem3Embedding, ImageEmbedding, TextRerank and their model catalogs) now sit behind it. Every ort-* feature, plus image-models, directml and webgpu, turns it on.
• qwen3 is split. qwen3-text is the text model and reranker, loading local files or GGUF, with no hf-hub and no image crates. qwen3-vl adds Qwen3VLEmbedding and needs hf-hub, since from_hf is its only loader. qwen3 is now both plus hf-hub, the same surface as before.
• A CI guard builds and tests qwen3-text with no ONNX Runtime, so a change that pulls ort back in fails the link.

The PR author reports 228 distinct crates before and 223 after for a qwen3,metal build, with no ort or ndarray left in it.

Do you need to change anything

Per the PR:

• Default builds: no. The dependency and feature trees are identical before and after.
• default-features = false with an ort-* feature or image-models: no.
• default-features = false, no ort feature, using the ONNX models (for example linking a system ONNX Runtime through ORT_LIB_LOCATION): add onnx, or the build breaks.

The rest of the day’s releases

• 7.4.0 also fuses Qwen3’s MLP activation into one Metal kernel and fixes Qwen3 vectors to be unit length in bf16.
• 7.4.1 scores the Qwen3 reranker in f32 without padding and tokenizes only the kept prefix of long Qwen3 texts.
• 7.5.0 runs the Qwen3 reranker’s shared prompt head once per score call.

The notes give no benchmark numbers for the Metal or reranker changes, so measure on your own hardware before you count on a speedup.