Tested models
The DVAI-Bridge Capacitor plugins are model-agnostic. Anything in GGUF form runs on capacitor-llama. Anything in MediaPipe .task form runs on capacitor-mediapipe. This page lists the specific models we exercise in CI and pre-release smoke tests, organized by tier of verification effort.
NOTE
Some entries below carry a placeholder tag — the upstream name wasn't finalized when this document was written. Verify the exact HF revision before pinning into your app.
Tier 1 — development (per-PR smoke)
Cached on the self-hosted CI runner. Loaded on every PR-level smoke run on iOS and Android.
| Backend | Model | Format | ~Size | Notes |
|---|---|---|---|---|
capacitor-llama | Llama-3.2-1B-Instruct | GGUF Q4_K_M | ~770 MB | Text completion baseline. contextSize: 2048, maxTokens: 256. |
capacitor-llama (embeddings) | bge-small-en-v1.5 | GGUF Q8_0 | ~133 MB | Started with embeddingMode: true. |
capacitor-mediapipe | gemma-3n-E2B-it (placeholder if gemma-4-E2B-it not yet shipped) | .task | ~1.5 GB | Vision-capable variant when the E2B .task artifact is published. |
capacitor-foundation | Apple-managed (implicit on iOS 26+) | — | 0 (zero-download) | No modelPath. Real-device required for full coverage; Simulator coverage is limited. |
Runtime memory is roughly 1.3–1.6× the on-disk size for GGUF — working buffers plus KV cache scaled by contextSize. .task artifacts are closer to 1.1–1.3× because MediaPipe manages memory differently.
Tier 2 — pre-release (manual, fuller coverage)
Run against a real device in the week before tagging a release.
| Backend | Model | Format | ~Size | Notes |
|---|---|---|---|---|
capacitor-llama | Qwen2.5-1.5B-Instruct (or Qwen3.5-1.5B once shipped) | GGUF Q4_K_M | ~1.0 GB | Multilingual smoke. |
capacitor-llama | Phi-3.5-mini-instruct | GGUF Q4_K_M | ~2.3 GB | Long-context (128k) smoke; cap to contextSize: 8192 on phones. |
capacitor-llama | Llama-3.2-3B-Instruct | GGUF Q4_K_M | ~2.0 GB | Larger-text-only smoke. |
capacitor-llama (vision) | gemma-4-E2B-it (placeholder) + matching mmproj | GGUF Q4_0 + mmproj | ~1.5 GB pair | Image content parts; requires mmprojPath. |
capacitor-llama (vision flagship) | gemma-4-E4B-it (placeholder) + mmproj | GGUF Q4_K_M + mmproj | ~3.5 GB pair | Higher-quality vision; phone-RAM-class. |
capacitor-llama (audio) | Phi-4-Multimodal-Instruct | GGUF Q4_K_M | ~7 GB | Audio + image; tablet / high-RAM phone only. |
capacitor-mediapipe | gemma-2b-it | .task | ~1.3 GB | Reference MediaPipe artifact from Google's tasks-genai distribution. |
Tier 3 — backend-specific real-device
| Backend | Notes |
|---|---|
capacitor-foundation | Apple's curated 3B-class model. Auto-loaded on iOS 26+. Cannot run on Simulator beyond a thin smoke; tier requires a real device with the system download present. |
capacitor-mediapipe (vision) | Vision-enabled Gemma .task variants from tasks-genai. Requires visionEnabled: true at session build-time; we wire this through StartOptions. |
Recommended start() parameters by class
| Model class | contextSize | gpuLayers | maxTokens |
|---|---|---|---|
| 1B text | 2048–4096 | 99 (full) | 256 |
| 1.5–3B text | 2048 | 99 | 256 |
| 1.5B vision (image) | 4096 | 99 | 384 |
| 3.5B vision (image) | 4096 | 99 | 384 |
| 7B multimodal (audio + image) | 4096 | 99 | 384 |
| Embeddings | 512 | 99 | n/a |
gpuLayers: 99 means "request maximum offload" — llama.cpp decides what fits. Lowering it manually is rarely needed. Raise the floor only if you see the device thermal-throttling under sustained inference.
What we deliberately don't verify
- Output quality. No BLEU, no benchmark accuracy. These tests check mechanics — load, respond, stream, free — not whether the model is good.
- Specific tokens-per-second figures. They vary too widely across device tiers for a per-PR gate.
- Long-running soak (>60s). Tier 2 covers a single round-trip per model. Longer sessions are an explicit user-test exercise pre-launch.
Curating this list
This file is the authoritative reference. Updates are docs-only commits. When upstream renames or removes a model, update both the entry above and the matching pinned URL and sha256 in your app's distribution config.