jobmatch-models

JobMatch on-device models

Weight bundle for the JobMatch iOS app. The app runs entirely on-device and fetches these files on first launch into its sandbox, mirroring this repo's directory layout exactly.

This repo is a distribution bundle, not a single model — it holds three separate models under three prefixes, each with its own provenance and licence.

Layout

llm/Qwen3.5-4B.Q4_K_M.gguf    # runs via llama.rn
embed/model.onnx              # onnxruntime
embed/tokenizer.json
embed/tokenizer_config.json
rerank/model.onnx             # onnxruntime
rerank/tokenizer.json
rerank/tokenizer_config.json

tokenizer.json and tokenizer_config.json are not interchangeable between the two encoders — they use different normalizers. Keep them with their model.

Versioning

The app pins a commit SHA, not main. Pushing here does not change what already-shipped installs download; a model update is a deliberate bump of the pinned SHA in the app's manifest.

Attribution

• Embedder — intfloat/multilingual-e5-large (MIT). The int8 ONNX export matches Xenova/multilingual-e5-large.
• Reranker — BAAI/bge-reranker-v2-m3 (Apache-2.0). Redistributed here in int8 ONNX form; see the upstream repo for the full licence text and NOTICE.
• LLM — Qwen/Qwen3.5-4B (Apache-2.0), quantized locally to GGUF.