feat(ai): ungate Gemma 3 4B and offer Q4_K_M / Q8_0 / f16 quants

[scotej]

Summary

• ggml-org/gemma-3-4b-it-GGUF is not gated (HF API gated: false, verified 2026-07-25) — drop the token requirement from the Best tier and update the stale keychain-help copy.
• Register the repo's other two quants (Q8_0, f16) as separate ModelSpec entries at the same pinned revision, sharing the f16 projector manifest. Each keeps its own install dir + benchmark record; existing gemma3-4b (Q4_K_M) installs keep their id and stay valid.
• New quantFamily grouping collapses the three into one picker card with a Quantization radio selector (busy variant pins the card; installed variants marked in the option label).
• ARCHITECTURE.md §8 table updated alongside the manifest (models.ts header rule). Retired the obsolete GatedNeedsToken story in favor of QuantVariantInstalled.

Manifest verification

Sizes + SHA256s taken from the HF API tree at pinned revision d09762... (LFS oids):

• gemma-3-4b-it-Q8_0.gguf 4 130 226 336 B
• gemma-3-4b-it-f16.gguf 7 767 474 336 B

Testing

• npm run build, lint, test (778 passed), check-tokens, check-strings, check-contrast all green.
• New unit tests: registry shape (6 specs / 4 tiers), no-gated invariant, Gemma family shares repo+revision+mmproj with distinct weight SHAs, modelCardGroups grouping.

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added selectable quantization variants for supported vision models, with download sizes and installed status shown in the model picker.
  • Grouped related model variants into a single card while preserving individual installation options.
  • Added additional Gemma 3 4B quantization options, including Q8 and F16.
• Documentation
  • Updated model guidance, download estimates, latency information, and Hugging Face token requirements.
  • Clarified model and projector pairing recommendations.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: c9a9db0a-7c9a-49bb-afbc-dd679c27ed88

📥 Commits

Reviewing files that changed from the base of the PR and between eb74005 and 12265ab.

📒 Files selected for processing (7)

• ARCHITECTURE.md
• src/features/ai/ModelPicker.tsx
• src/features/ai/index.ts
• src/features/ai/models.ts
• src/stories/ModelPicker.stories.tsx
• src/strings.ts
• tests/unit/ai-models.test.ts

📜 Recent review details

⏰ Context from checks skipped due to timeout. (2)

• GitHub Check: Rust (Windows)
• GitHub Check: Frontend

🧰 Additional context used

📓 Path-based instructions (9)

src/**/*.{ts,tsx,css}

📄 CodeRabbit inference engine (CLAUDE.md)

src/**/*.{ts,tsx,css}: Use design tokens instead of hard-coded visual values in frontend TypeScript, TSX, and CSS files.
Meet WCAG AA requirements for every text/background pairing in both themes; do not convey information by color alone.
Respect reduced-motion preferences through the global kill switch; new motion sites must be gated by default.

Files:

• src/features/ai/index.ts
• src/strings.ts
• src/stories/ModelPicker.stories.tsx
• src/features/ai/models.ts
• src/features/ai/ModelPicker.tsx

src/**/*.{ts,tsx}

📄 CodeRabbit inference engine (CLAUDE.md)

src/**/*.{ts,tsx}: Keep JSX text and aria-label literals consistent with the voice defined in DESIGN-SYSTEM.md §14, and prefer centralized strings.
Treat peer wire formats and identity derivation as cross-version contracts; coordinate changes so older peers and existing stored data remain compatible.
Do not add telemetry; the application must remain local-only.

Files:

• src/features/ai/index.ts
• src/strings.ts
• src/stories/ModelPicker.stories.tsx
• src/features/ai/models.ts
• src/features/ai/ModelPicker.tsx

src/**/*.{ts,tsx,rs}

📄 CodeRabbit inference engine (CLAUDE.md)

Never instruct users to paste a model file or BIP39 mnemonic into an AI service or chat.

Files:

• src/features/ai/index.ts
• src/strings.ts
• src/stories/ModelPicker.stories.tsx
• src/features/ai/models.ts
• src/features/ai/ModelPicker.tsx

**/*.{ts,tsx,rs}

📄 CodeRabbit inference engine (CLAUDE.md)

**/*.{ts,tsx,rs}: Add comments only when the reason is non-obvious; identifiers should carry meaning and code should read top-to-bottom.
Maintain scope discipline: do not refactor adjacent code during feature work or add abstractions for hypothetical future needs.

Files:

• src/features/ai/index.ts
• tests/unit/ai-models.test.ts
• src/strings.ts
• src/stories/ModelPicker.stories.tsx
• src/features/ai/models.ts
• src/features/ai/ModelPicker.tsx

**/*.{ts,tsx}

📄 CodeRabbit inference engine (CLAUDE.md)

Use TypeScript strict mode and ensure frontend code passes the TypeScript build and type checks.

Files:

• src/features/ai/index.ts
• tests/unit/ai-models.test.ts
• src/strings.ts
• src/stories/ModelPicker.stories.tsx
• src/features/ai/models.ts
• src/features/ai/ModelPicker.tsx

**/*

📄 CodeRabbit inference engine (CLAUDE.md)

**/*: Before committing or opening a PR, run the required build, lint, test, design-token, strings, contrast, accessibility, formatting, and relevant Rust quality gates.
Use one focused change per commit with a Conventional Commit subject such as feat:, fix:, chore:, docs:, or ci:; PRs should squash-merge.

Files:

• src/features/ai/index.ts
• ARCHITECTURE.md
• tests/unit/ai-models.test.ts
• src/strings.ts
• src/stories/ModelPicker.stories.tsx
• src/features/ai/models.ts
• src/features/ai/ModelPicker.tsx

**/*.test.ts

📄 CodeRabbit inference engine (CLAUDE.md)

Write Vitest tests for applicable unit and integration behavior; component tests are not currently supported without an explicit test-harness scope expansion.

Files:

• tests/unit/ai-models.test.ts

src/strings.ts

📄 CodeRabbit inference engine (CLAUDE.md)

Keep toast and notification copy in src/strings.ts; prefer this module for user-facing strings.

Files:

• src/strings.ts

src/**/*.stories.{ts,tsx}

📄 CodeRabbit inference engine (CLAUDE.md)

src/**/*.stories.{ts,tsx}: Provide Storybook coverage for every primitive and feature component.
Ensure every Storybook story passes the axe-core accessibility gate.

Files:

• src/stories/ModelPicker.stories.tsx

🔇 Additional comments (7)

src/features/ai/models.ts (1)

49-53: LGTM!

Also applies to: 125-204, 238-270

src/features/ai/index.ts (1)

12-14: LGTM!

tests/unit/ai-models.test.ts (1)

10-10: LGTM!

Also applies to: 47-93

ARCHITECTURE.md (1)

448-451: LGTM!

src/features/ai/ModelPicker.tsx (1)

2-5: LGTM!

Also applies to: 28-35, 201-217, 231-259, 328-362

src/strings.ts (1)

1151-1151: LGTM!

Also applies to: 1466-1472

src/stories/ModelPicker.stories.tsx (1)

215-229: LGTM!

---

📝 Walkthrough

Walkthrough

Gemma 3 4B now has multiple registered quant variants sharing a projector, grouped into one model card with quant selection. Picker copy, Storybook coverage, architecture documentation, exports, and registry tests were updated accordingly.

Changes

Vision model quantization

Layer / File(s)	Summary
Model registry and quant-family grouping src/features/ai/models.ts, src/features/ai/index.ts, tests/unit/ai-models.test.ts, ARCHITECTURE.md	ModelSpec supports quantFamily; Gemma variants share projector metadata while retaining distinct model files, and modelCardGroups groups variants into picker entries. Registry expectations and architecture notes were updated.
Quantized model card selection src/features/ai/ModelPicker.tsx, src/strings.ts, src/stories/ModelPicker.stories.tsx	Model cards render grouped variants, select the active quant based on busy/chosen/installed state, and display quant labels, sizes, and installed status. Copy and Storybook coverage were updated.

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

Sequence Diagram(s)

sequenceDiagram
  participant ModelPicker
  participant modelCardGroups
  participant ModelCard
  participant PickerState
  ModelPicker->>modelCardGroups: group SUPPORTED_MODELS by quantFamily
  modelCardGroups-->>ModelPicker: return variant groups
  ModelPicker->>ModelCard: pass per-variant PickerState entries
  ModelCard->>PickerState: select busy, chosen, installed, or first variant
  ModelCard-->>ModelPicker: render selected quantization actions

Loading

Suggested reviewers: copilot, claude

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: ungating Gemma 3 4B and adding multiple quant variants.
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.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

✨ Simplify code

• Create PR with simplified code
• Commit simplified code in branch fix/gemma-ungated-quants

---

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.}