Add Gemma 4 and Qwen3.6/Qwen3.8 catalog rows

[sanchitmonga22]

Summary

Catalog-only change to RunAnywhereAI/Core/Services/ModelCatalogBootstrap.swift, added over two rounds on this branch. No SDK/Package.swift/dependency changes, no new native code — pure registerLLM/registerMultiFile rows following the file's existing conventions.

Round 1 — Gemma 4 and Qwen3.6/Qwen3.8

GGUF rows (framework .llamaCpp, text-only chat models)

id	model	quant	file size
gemma-4-e2b-it-q4_k_m	Gemma 4 E2B IT	Q4_K_M	3.11 GB
gemma-4-e4b-it-text-q4_k_m	Gemma 4 E4B IT	Q4_K_M	4.98 GB
gemma-4-12b-it-q4_k_m	Gemma 4 12B IT	Q4_K_M	7.12 GB
gemma-4-26b-a4b-it-q4_k_xl	Gemma 4 26B-A4B IT (MoE, 26B total/4B active)	UD-Q4_K_XL	17.01 GB
gemma-4-31b-it-q4_k_m	Gemma 4 31B IT (dense)	Q4_K_M	18.32 GB
gemma-4-31b-it-q2_k_xl	Gemma 4 31B IT (dense)	UD-Q2_K_XL	11.77 GB
qwen3.8-27b-q4_k_m	Qwen3.8 27B (dense)	Q4_K_M	17.11 GB
qwen3.6-35b-a3b-q4_k_m	Qwen3.6 35B-A3B (MoE, 35B total/3B active)	UD-Q4_K_M	22.13 GB

All from the plain unsloth/...-GGUF repos, matching the existing Qwen3.x rows' source convention. Qwen rows set supportsThinking: true.

Gemma 4 E4B id note: this file already had a multimodal gemma-4-e4b-it-q4_k_m row (a VLM decoder+mmproj pair from ggml-org). To avoid an id collision with a materially different artifact, the new text-only row uses gemma-4-e4b-it-text-q4_k_m instead.

MLX rows added (framework .mlx)

id	model	size
mlx-qwen3.8-27b-4bit	Qwen3.8 27B 4bit	~16.05 GB
mlx-qwen3.6-35b-a3b-4bit	Qwen3.6 35B-A3B 4bit	~20.40 GB

Reuse the file's existing MLX pattern exactly: the mlxCatalogEnabled gate lives inside the shared registerLLM helper (guard framework != .mlx || mlxCatalogEnabled else { return }) — no new plumbing invented.

MLX Gemma 4 rows intentionally NOT added, at any size. This file already documented that mlx-swift-lm 3.31.4's Gemma4TextAttention aborts on Gemma 4's asymmetric QK-norm. That bug is architectural, not per-size, so it blocks 12B/26B-A4B/31B exactly as it already blocked E2B/E4B — I extended the existing comment to say so explicitly.

Round 2 — Muse Glimmer 30B, IBM Granite 4.1, Nemotron Nano Omni (deep HF verification pass)

New GGUF rows

id	model	quant	notes
muse-glimmer-30b-q4_k_xl	Meta Muse Glimmer 30B (VLM, mmproj)	UD-Q4_K_XL	Apache 2.0, released 2026-08-10
granite-4.1-3b-q4_k_m	IBM Granite 4.1 3B	Q4_K_M	Apache 2.0, dense
granite-4.1-8b-q4_k_m	IBM Granite 4.1 8B	Q4_K_M	Apache 2.0, dense
granite-4.1-30b-q4_k_m	IBM Granite 4.1 30B	Q4_K_M	Apache 2.0, dense (verified: GGUF metadata shows plain granite arch, no expert-routing fields)
nemotron-3-nano-omni-30b-a3b-reasoning-q4_k_m	NVIDIA Nemotron 3 Nano Omni 30B-A3B Reasoning (VLM, mmproj)	UD-Q4_K_M	MoE, 31B total/3B active; supportsThinking: true

New MLX rows

id	model
mlx-granite-4.1-3b-4bit	MLX IBM Granite 4.1 3B 4bit
mlx-granite-4.1-30b-4bit	MLX IBM Granite 4.1 30B 4bit

No MLX row for Granite 4.1 8B: mlx-community only publishes that size in bf16/nvfp4/mxfp4/mxfp8 (no clean 4-bit from a reputable quantizer) plus one unofficial third-party "-oQ4" repo, which was excluded per the "stick to reputable quants" rule already used for Gemma/Qwen.

MLX rows investigated and NOT added, with verified reasons

I checked out the exact pinned mlx-swift-lm 3.31.4 source (from Package.resolved, matched against the local SwiftPM checkout) rather than assuming compatibility:

• mlx-community/Muse-Glimmer-30B-4bit — its config.json declares model_type: "muse_glimmer". VLMModelFactory's registered types (paligemma, qwen2_vl, qwen2_5_vl, qwen3_vl, qwen3_5[_moe], idefics3, gemma3, gemma4[_unified], smolvlm, fastvlm, llava_qwen2, pixtral, mistral3, lfm2_vl, glm_ocr) contain no muse_glimmer entry — this would download in full and then fail to load. Same class of bug as the existing documented Gemma 4 MLX note, so I extended that note rather than silently omitting the row.
  • Separately, I caught a real data error: the originally supplied size estimate for this repo (~19.4 GB) does not match the live HF blobs API, which reports 4 safetensors shards totaling 39,443,521,145 B (~39.4 GB) — roughly 2x. Flagged in the code comment either way, since the row isn't being added.
• mlx-community/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-4bit — config.json declares model_type: "NemotronH_Nano_Omni_Reasoning_V3". VLMModelFactory has zero Nemotron entries of any kind; LLMModelFactory's only Nemotron entry, "nemotron_h", is a different, non-matching string and is text-only regardless (no vision path). Not added.
• IBM Granite 4.1 (3B/30B) — by contrast, both repos' config.json declare plain model_type: "granite" (GraniteForCausalLM), which is a registered LLMModelFactory entry (GraniteConfiguration/GraniteModel). Added.

STT/TTS investigated and NOT added — SDK version verification

Checked runanywhere-sdks core/VERSIONS at the exact tag (v0.20.19) this app's Package.resolved pins: SHERPA_ONNX_VERSION_IOS/_MACOS = 1.13.2.

• NVIDIA Nemotron-3.5-ASR-Streaming 0.6B — needs sherpa-onnx v1.13.5 (2026-08-11) for native streaming-transducer support. Vendored is 1.13.2. Not added — blocked on the SDK bumping its sherpa-onnx pin, not a dead-end.
• Supertone Supertonic v3 TTS — vendored 1.13.2 does add native C++ Supertonic support (confirmed: offline-tts-supertonic-*.{h,cc} exist in the vendored sherpa-onnx source). But the published Supertone/supertonic-3 HF repo ships onnx/unicode_indexer.json and ten per-speaker voice_styles/{F,M}[1-5].json files, while the vendored loader's OfflineTtsSupertonicModelConfig hard-requires a single binary unicode_indexer.bin (asserts the path ends in .bin) and a single combined voice.bin blob, both produced by an offline conversion script (generate_indexer_bin.py) that ships with neither the HF repo nor this codebase. I also found zero references to "supertonic" anywhere in this SDK's own engine/provider code, so the app-side dispatch to this loader doesn't exist yet either. Not added — two real blockers, not one; needs a RunAnywhere-side re-export (the same pattern already used for Parakeet CTC) plus provider wiring.

Explicitly considered and NOT added (all rounds)

• NVFP4 quants (e.g. unsloth/Qwen3.8-27B-NVFP4) — NVIDIA-specific; neither of this app's backends can load it.
• Abliterated/uncensored/heretic/personality-modified community quants — only plain unsloth/...-GGUF and mlx-community/...-4bit repos were used anywhere in this PR.
• Jina Reranker v3.5 — real, but CC-BY-NC-4.0 non-commercial; needs product/legal sign-off before bundling in a commercial app.
• IBM Granite Speech 4.1 2B — real, but an audio-instruction encoder+decoder-LM, architecturally different from the ASR transducer/CTC models this app's STT path expects.
• IBM Granite Vision 4.1 4B — no unsloth/mlx-community/reputable quant exists yet.
• Ultra-large frontier releases (Kimi K3, GLM-5.2, DeepSeek V4, Mistral Large 3, MiniMax M3) — real, but hundreds-of-GB-to-TB even at 4-bit; not storable/loadable on any of this app's targets.

No license field exists anywhere in this file's registerLLM/registerMultiFile API (grepped — zero hits across the whole file both rounds), so none was added to any new row; matches every existing row.

Verification (both rounds)

• swift package resolve — clean.
• ./scripts/smoke.sh — passes.
• ./scripts/verify.sh (real xcodebuild against iphonesimulator) — BUILD SUCCEEDED, no warnings on the changed file, after both commits.
• No duplicate id: values anywhere in ModelCatalogBootstrap.swift after either round (checked with a full-file dedupe pass).

🤖 Generated with Claude Code

https://claude.ai/code/session_01DxGfpsMbkNzoQT559b2tXz

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: b08b3f6a-2e8b-4dc2-a541-1e1f29f96aac

📥 Commits

Reviewing files that changed from the base of the PR and between 939c244 and 3ba8f10.

📒 Files selected for processing (1)

• RunAnywhereAI/Core/Services/ModelCatalogBootstrap.swift

Included review availability: Your plan includes up to 1 review per rolling hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

The model catalog adds Qwen3 and Gemma 4 variants for Llama.cpp. It adds Qwen3 variants for MLX and documents why MLX Gemma 4 variants remain unregistered.

Changes

Model catalog expansion

Layer / File(s)	Summary
Llama.cpp model registrations RunAnywhereAI/Core/Services/ModelCatalogBootstrap.swift	Adds Qwen3.8 27B, Qwen3.6 35B-A3B, and five text-only Gemma 4 registrations with identifiers, URLs, quantizations, memory requirements, and supported thinking settings.
MLX model registrations and limitation RunAnywhereAI/Core/Services/ModelCatalogBootstrap.swift	Adds 4-bit MLX entries for the two Qwen3 models. The Gemma 4 exclusion note now covers all sizes and cites unavailable k_norm weights required by the current loader.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Merge Risk: ⚪ Minimal · up to 3ba8f

This change adds model catalog entries without altering SDK dependencies or native runtime code, and the reported checks pass. No actionable merge-blocking risk remains beyond normal review.

Possibly related PRs

• RunanywhereAI/runanywhere-ios#4: Both changes update ModelCatalogBootstrap.swift with model catalog entries.

Suggested reviewers: ayaangazali, siddhesh2377

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change by naming the added Gemma 4 and Qwen3.6/Qwen3.8 catalog rows.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch add-gemma4-qwen3.6-qwen3.8-models

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}