econetworkai/Qwen3-0.6B-ONNX-external-data

Qwen3-0.6B ONNX q4f16 — external-data repack (Eco)

This repository hosts a byte-preserving repack of the q4f16 ONNX build of Qwen3-0.6B for Eco, the local-first in-browser AI chat app. The single file onnx/model_q4f16.onnx from the upstream repository is split into a graph file plus an external-data companion:

file	bytes	sha256
onnx/model_q4f16.onnx	328,247	c4bb4067156a97c57bc92d945f538dfc7db92153835e9c1f00e598b848f66411
onnx/model_q4f16.onnx_data	569,493,504	1ec7609fa8fbec10c830c440087acddfba3d8d5204d457cb8061b732e3a137d8

Why

onnxruntime-web stages a single-file model wholesale inside the WebAssembly heap while building the inference session, and frees that staging copy only afterwards — so browsers pay a transient of roughly one extra copy of the weights at load time. Files supplied as externalData are mounted outside the heap and skip the staging copy. Measured on this model (paired cells, 2026-07-16): load-time memory plateau reduced ~600–700 MB, permanent WASM reservation reduced ~789 MB, decode speed unchanged.

Provenance (verify it yourself)

• Upstream: onnx-community/Qwen3-0.6B-ONNX at revision da1453100cf3ff33ef56d17983fc7a8648706db6 (onnx/model_q4f16.onnx, sha256 9e33a5911974174761d0dfdcc0bec975d9c45af0eae5e9eb647b8ba9442a8f91).
• Transformation: tensor bytes are moved, not modified — the ONNX graph references the same data at offsets in the .onnx_data file. Produced by the deterministic script scripts/repack-onnx-external-data.py with onnx==1.22.0; re-running it on the upstream file reproduces the sha256s above byte-for-byte.
• Tokenizer/config files are verbatim copies from the upstream repository at the same revision (so this repository is a self-contained artifact pin); their content hashes are unchanged from upstream.

License and attribution

Qwen3 is released by the Qwen team (Alibaba Cloud) under Apache-2.0; the ONNX conversion is by the onnx-community maintainers. This repack preserves their license and adds no weight changes.