[iscekic]

Summary

• Replace the curated Llama 3.2 1B Instruct entry with Qwen3 0.6B (Q8_0) from the official Qwen/Qwen3-0.6B-GGUF repository.
• The Llama 3.2 Community License requires a copy of the agreement and a "Built with Llama" notice when a product makes the model available. The app showed only the license name.
• Qwen3 0.6B uses Apache 2.0, like the other curated entries. The file is smaller than the Llama file and fits the same slot between Qwen2.5 0.5B and 1.5B.
• The entry pins revision 23749fefcc72300e3a2ad315e1317431b06b590a with the size and SHA-256 that Hugging Face publishes for it.
• Downloaded models keep their own record, so a device that already downloaded Llama keeps it. The catalog no longer offers it.

Verification

• vitest run src/lib/chat/gguf-catalog.test.ts src/lib/chat/hf-catalog.test.ts passes.
• A HEAD request on the pinned resolve URL returns x-linked-size and x-linked-etag that match the entry.
• The vendored llama.cpp in llama.rn 0.12.9 (build 6c8dcaa) knows the qwen3 architecture. The model chat template reads enable_thinking, so the app offers its thinking switch.
• On a VM, the downloaded file matched the pinned size and SHA-256.
• On the same VM, llama-server built from 6c8dcaa loaded the file with --jinja --reasoning-format auto. It answered with thinking off, answered with separate reasoning with thinking on, and returned a get_weather tool call.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Executive Summary

The change is a self-contained data substitution in the curated GGUF catalog; the new entry is internally consistent and no downstream consumer keys on the removed fileId.

Files Reviewed (1 file)

• apps/mobile/src/lib/chat/gguf-catalog.ts

Notes

• New entry is well-formed: revision 23749fefcc72300e3a2ad315e1317431b06b590a is 40-hex, sha256 is 64-hex, sizeBytes is positive, and license Apache 2.0 matches the sibling curated entries.
• No code references the removed llama-3.2-1b-instruct-q4_k_m fileId; GGUF_CATALOG consumers (model-catalog-screen.tsx, hf-catalog.ts CURATED keyed by sha256, gguf-records.ts) are records self-contained per entry, so removing a catalog entry does not affect already-downloaded models or cleanup.
• Could not independently verify the pinned file's sizeBytes/sha256 against Hugging Face in this read-only sandbox; the PR body states these were validated via a HEAD request and a load test. This is the only unverified aspect.

---

_{Reviewed by deepseek-v4.1-flash · Input: 73.4K · Output: 5.1K · Cached: 479K}

_{Review guidance: REVIEW.md from base branch main}