Expose Qwen 3.5, Qwen 3 - 4B and Phi-4-mini from the executorch registry

[kfaracik]

Part of the multilingual audit. Child of #287.

Problem

react-native-executorch already exports these models, and DEFAULT_MODELS (constants/default-models.ts) does not list any of them:

Export	Why it belongs in the hub
QWEN3_5_0_8B_QUANTIZED	newest Qwen generation, phone-sized
QWEN3_5_2B_QUANTIZED	same generation at the Gemma 4-2B quality tier
QWEN3_4B_QUANTIZED	for 8 GB+ devices
PHI_4_MINI_4B_QUANTIZED	broad declared multilingual coverage, incl. Polish and Turkish
LFM2_5_350M	smallest fallback

Qwen is the broadest multilingual family available to us — Qwen 3 declares 119 languages including hi, ur, ar, fa, id, tr and pl — and Qwen 3.5 is a generation newer than anything currently shipped.

This is plumbing: an entry in DEFAULT_MODELS per model, with parameters, modelSize and labels. No new dependency, no new download infrastructure, no native work. Probably the cheapest single improvement available for non-English users.

What to do

1. Add the entries. Sizes must be measured, not guessed — modelSize drives the download-size UI and, on Gemma, already differs per platform.
2. Verify each download URL actually resolves on both platforms before shipping. There is precedent for a stale weight path in this registry, and a 404 in the Model Hub is a dead end for the user.
3. Check the iOS backend per model. Gemma 4 - 2B is a known trap: its MLX build is gated off the A15, so it will not download on a 4 GB iPhone SE even though the URL is valid. That is a device-backend restriction, not a broken link, and it needs handling as such rather than surfacing as a failed download.
4. Run the multilingual sweep (Restore the on-device eval harness and extend the corpus to the actual top languages #282) on the new models before recommending them anywhere.

Note on the Qwen 3.5 claim

Its multilingual coverage is asserted here on family inheritance, not on a read model card. Read the card before publishing any user-facing language claim.

Related

• Allow importing model directly from HuggingFace #168 — importing models directly from HuggingFace, a more general solution to the same "not enough models" problem. This issue is the cheap version and does not block that one.