| Model | Vendor | Params | Size | Context | Modality | License | Min RAM | In Cove? | Details |
|---|---|---|---|---|---|---|---|---|---|
| Gemma 4 E2B | Google DeepMind | 2.3B | 1.5 GB | 128,000 | text+vision+audio | apache-2.0 | 4 GB | ✓ | View → |
| Microsoft Phi-4 multimodal | Microsoft Research | 5.6B | 3.5 GB | 128,000 | text+vision+audio | mit | 6 GB | ✓ | View → |
| Apple Foundation Models | Apple | 3B | Not disclosed | Not disclosed | text+vision | apple-proprietary | 8 GB | ✓ | View → |
| Llama 3.2 Mobile | Meta AI | 3B | 2 GB | 128,000 | text | llama-community | 6 GB | ✓ | View → |
| Qwen 3.5 2B | Alibaba Cloud | 2B | 1.5 GB | 262,000 | text+vision | apache-2.0 | 4 GB | ✓ | View → |
| Ministral 3B | Mistral AI | 3B | 2 GB | 32,768 | text+vision | apache-2.0 | 4 GB | ✓ | View → |
| DeepSeek R1 Distill (Qwen 1.5B) | DeepSeek | 1.5B | 1 GB | 32,768 | text | apache-2.0 | 4 GB | ✓ | View → |
| MiniCPM-V 4.0 | ModelBest / OpenBMB | 4.1B | 2.5 GB | 32,768 | text+vision | modelbest-terms | 4 GB | ✓ | View → |
How we built this leaderboard. All 8 models are evaluated against the same dimensions — parameters, quantized size, context window, modality, license, and minimum device RAM — sourced from official model cards (Hugging Face, vendor blogs, official documentation) as of the last-reviewed date shown above. We do not run our own benchmarks; instead, we cross-reference 2-3 authoritative sources per data point and prefer the vendor's own claim where it conflicts with third-party reproduction. Numbers may diverge from your real-world experience by ±10-20% depending on quantization scheme (Q4_K_M, AWQ, GPTQ all behave differently), runtime (LiteRT, MediaPipe, ExecuTorch, llama.cpp, Core ML), and device thermal throttling. Each model card carries its own `lastReviewed` field; this page is refreshed every quarter. Conflicts and ambiguities are tracked in our open GitHub repo.
Run the largest mobile-optimized models comfortably
Sweet spot for size/capability balance
Smaller models that still deliver real value
