jinaai/jina-embeddings-v4 config.json

JinaEmbeddingsV4Model, Qwen2.5-VL text backbone, hidden_size 2048, 36 layers, 16 heads, 2 KV heads, max_position_embeddings 128000, sliding_window 32768, multi_vector_projector_dim 128, matryoshka_dims [128,256,512,1024,2048], vocab 151936. Primary source for architecture fields. W10 LOWCONFIDENCE PROVENANCE AUDIT 2026-10-10: re-classified UNCHECKED after an actual re-fetch. Fetched over HTTPS from this host: HTTP 200, the returned file is live JSON and parses. Every architecture field this note names was then read back out of the live config.json and matches the note's value field-for-field. The 0.5 rating therefore marked a lookup nobody had run, not a property of the source: it is the model publisher's own repository, so it is a primary-shaped source, and the one caveat is that it is a mirror for the license-gated originals (see 'a-hosting-platform-is-not-a-vendor' and the gated-repo records). Raised 0.5 -> 0.88, status draft -> verified.