You can deploy this model and start earning money today!

weave-eng/
qwen3-embedding-0.6b-onnx-router

Name: qwen3-embedding-0.6b-onnx-router
Availability: InStock

onnx

embeddings

weave-router

Qwen/Qwen3-Embedding-0.6B

quantized

apache-2.0

Views

No views yet

Qwen3-Embedding-0.6B ONNX export for the Weave router

INT8 ONNX export of Qwen/Qwen3-Embedding-0.6B (Apache-2.0) used by the Weave router's cluster scorer. Produced by scripts/export_qwen3_onnx.py in the workweave/router repo.

Files

File	Purpose
`model.onnx`	INT8 dynamic quantization, deployed artifact
`model_fp32.onnx` + `.data`	fp32 export, provenance only (not deployed)
`tokenizer.json`, `tokenizer_config.json`	HF tokenizer
`fixture.json`	Go parity-test fixture (mirrors `internal/router/cluster/testdata/fixture_qwen3.json`)

Export details

Pooling baked into the graph: attention-mask-aware last-token pooling; output is sentence_embedding [batch, 1024]. No L2 norm in the graph (the router's hugot pipeline applies it).
Opset 18, exported with the torch dynamo exporter.
Quantization: INT8 dynamic, per-channel, with down_proj (3072x1024) and o_proj (2048x1024) MatMuls kept fp32 (outlier-sensitive activations).
Parity (cosine vs sentence-transformers reference, no instruction prefix): fp32 worst 0.99985, INT8 worst 0.98311. Go-vs-Python worst 0.9808.

Versioning

The router Dockerfile pins HF_QWEN_REVISION to a commit SHA of this repo. Revisions are immutable; pushing a new export here never changes deployed images until the pin is bumped deliberately.

weave-eng/qwen3-embedding-0.6b-onnx-router

Qwen3-Embedding-0.6B ONNX export for the Weave router

Files

Export details

Versioning

weave-eng/
qwen3-embedding-0.6b-onnx-router