feat: add PARAPHRASE_MULTILINGUAL_MINILM_L12_V2 text embeddings model

[msluszniak]

Description

Adds the paraphrase-multilingual-MiniLM-L12-v2 sentence-transformer model — the second multilingual embeddings model after distiluse, completing #945. Ships only the XNNPACK 8da4w variant under MODEL_REGISTRY.ALL_MODELS (see "Why a single variant" below).

384-d output, max 126 tokens, 50+ languages. Tokenizer is Unigram + Precompiled normalizer + Metaspace decoder — requires the bumped pytorch/extension/llm/tokenizers runtime from #1114, so this PR blocks on that landing first and should be rebased onto main once #1114 merges.

HF repo: software-mansion/react-native-executorch-paraphrase-multilingual-MiniLM-L12-v2 (v0.9.0 tag, layout mirrors distiluse).

Why a single variant —
TLDR 8da4w works faster then all and was also one of the smallest, without loss in precision.
Longer answer:
unlike distiluse, where Core ML fp32 won iPhone thanks to ANE acceleration, benchmarks on iPhone 17 Pro + OnePlus 12 (~80-token input, 50 measured forwards after 3 warmups) showed the XNNPACK 8da4w variant Pareto-dominates the other three on both platforms: faster than XNNPACK fp32, Core ML fp32 and Core ML fp16 on iPhone, and ~36% smaller steady-state memory footprint than the next-best variant. Likely cause: paraphrase-multilingual-MiniLM-L12-v2 is a smaller model (~118 M params, 12 layers) where Core ML's runtime doesn't push enough work onto ANE for the precision-conversion overhead to pay off. fp16 being slower than fp32 on Core ML for this model is a tell that the runtime is falling back to slower compute units. Shipping only _8DA4W keeps the public surface aligned with the data; if a future Core ML or model update flips the verdict, easy to add the other variants back.

Memory methodology note — the new paraphrase row in docs/docs/02-benchmarks/memory-usage.md reports RSS / phys_footprint deltas from a clean app baseline (loaded − idle), captured on-device at the same conceptual point. The existing distiluse rows there (36 / 44 MB) come from an older measurement pass with a different (and not reconstructable from the diff) methodology, so the two rows are not directly comparable. A separate pass to re-measure distiluse and other rows with the same methodology would be a good follow-up.

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Other (chores, tests, code style improvements etc.)

Tested on

• iOS
• Android

Testing instructions

1. cd apps/text-embeddings && npx expo run:ios (or run:android).
2. Pick "Multilingual Paraphrase (8da4w)" in the model picker.
3. Add a sentence in one language, query with an aligned sentence in another (e.g. Polish "Słoneczko" against "It's so sunny outside!"). The cross-lingual pair should top the matches.

Related issues

Closes the paraphrase-multilingual half of #945 (the distiluse half landed in #1098).

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes

Blocks on #1114.

[msluszniak]

Would you rather keep 8da4w suffix or you are in favour of dropping it? cc: @chmjkb @NorbertKlockiewicz @barhanc

[barhanc]

Model works, just this small comment regarding naming.

[barhanc]

I think for now it should be PARAPHRASE_MULTILINGUAL_MINILM_L12_V2_QUANTIZED so that we maintain consistent naming (and we should probably retroactively change the DISTILUSE_BASE_MULTILINGUAL_CASED_V2_8DA4W also). We had some discussion about model registry and naming (#612) so we can change the symbols all in one go when implementing that.

[chmjkb]

agreed, I think that answers your question @msluszniak

[msluszniak]

Ok, so I'll change the name for this model and for distiluse we will fix this in one holistic change.