Vamp Assistant 0.10.48 — managed inference runtimes

iOS distribution update — October 1, 2026: Public IPA downloads have been retired as distribution moves to the App Store. Official App Store links will be shared when available. Historical release notes and source code remain available.

Vamp Assistant 0.10.48 (127) adds optional, separately installed inference runtimes while preserving the existing GGUF and MLX paths.

• Managed oMLX can reuse prompt caches for verified small MLX models. App-level text, image, unload and fallback checks passed. The Qwen3 0.6B catalog entry offers a small text checkpoint for this path; it is less capable than the larger models.
• Native MLXFast Bonsai and a pinned newer llama.cpp Metal runtime are available as experiments. Both are disabled below 24 GB after longer tests exceeded the 16 GB M4's memory budget. They are not claimed as working speedups on that Mac. Existing Huihui and Bonsai GGUF runtimes are preserved.
• A harness that exits during startup now reports its error instead of crashing the app with SIGPIPE.
• iPhone/iPad retains Photos/Files image attachments, persistent image drafts and queued images, the manual queued-message send action, unload control, and the simpler new-session flow.

Runtimes are installed separately with scripts/install-inference-runtimes.py; they are not bundled in the downloads. Settings remain off by default. See docs/M4-PERFORMANCE-REVIEW.md for pins, measured results, limits and setup.

Distribution: the Apple Silicon Mac DMG is a development-signed preview, not notarized. The iOS IPA is unsigned for AltStore/SideStore/Sideloadly or your own signing workflow.

Validation: 1,059 macOS tests passed (47 opt-in skips), 44 iOS unit/UI tests passed, and live managed oMLX text/image/unload and short Huihui Metal vision/unload gates passed. Longer Metal and MLXFast tests exposed memory limits; those backends remain disabled on 16 GB Macs.