1of2/jina-code-embeddings-0.5b-coreml-w16

jina-code-embeddings-0.5b — Core ML bundle

On-device Core ML conversion of jinaai/jina-code-embeddings-0.5b, produced by the Glossematics conversion pipeline. The repository root is a valid model bundle (manifest.json, formatVersion 2) consumed directly by the Glossematics Swift package — download the snapshot and point GlossModelBundle at it.

• Embedding dimension: 896 (Matryoshka: [64, 128, 256, 512, 896])
• Towers: text
• Text buckets: [32, 64, 128, 256, 512, 1024, 2048] + batch-4 functions at [32, 64, 128]
• Bidirectional text tower with explicit attention mask: False
• Artifacts: precompiled .mlmodelc (loads directly, no on-device compile)
• Numeric recipe: float16 weights / float16 activations — embedding space jinaai/jina-code-embeddings-0.5b:896:w16a16

Usage (Swift, macOS 15 / iOS 18)

import Glossematics

let bundle = try await GlossModelBundle(hub: "1of2/jina-code-embeddings-0.5b-coreml-w16")
let gloss = GlossOmniEmbedder(bundle: bundle)
let q = try await gloss.embed(text: "What is the capital of France?", prompt: .query)

License

The source model weights are CC-BY-NC-4.0 (© Jina AI) and this converted artifact is a derivative work distributed under the same license: non-commercial use only, with attribution. Commercial use requires a license from Jina AI. See the source model card for details.

Provenance & validation

• Source: jinaai/jina-code-embeddings-0.5b revision 4db235132dafbe56a8b9c5f59b59795ecf58a4a7
• Converter: jina_convert 0.2.0 (commit 4fd131b)
• Validated against fp32 references captured from the source model (per-modality cosine parity gates; end-to-end through the public Swift API via gloss-bundle-smoke).