Model catalog
On-device AI model catalog
This is a source-linked catalog, not a leaderboard. Cove has not run a common benchmark for these entries. A model card describes a model; it does not prove that a particular phone, runtime, or app can run it.
| Model | Publisher | Parameters | Configured context | Documented input | License | Publisher source |
|---|---|---|---|---|---|---|
| Gemma 4 E2B Gemma 4 E2B-it | Google DeepMind | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Publisher source ↗ |
| Microsoft Phi-4 multimodal Phi-4-multimodal | Microsoft Research | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Publisher source ↗ |
| Apple Foundation Models SystemLanguageModel (OS-managed) | Apple | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Publisher source ↗ |
| Llama 3.2 1B / 3B Llama 3.2 1B / 3B | Meta AI | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Publisher source ↗ |
| Qwen 3.5 2B Qwen3.5-2B | Alibaba Cloud | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Publisher source ↗ |
| Ministral 3 Ministral 3 v25.12 | Mistral AI | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Unverified in this catalog | Publisher source ↗ |
| DeepSeek R1 Distill (Qwen 1.5B) DeepSeek-R1-Distill-Qwen-1.5B | DeepSeek | 1.5B | 131,072 configured positions | Text | Card lists MIT; its notes identify Qwen 2.5 base weights as originally Apache 2.0. Check terms for the exact artifact. | Publisher source ↗ |
| MiniCPM-V 4.0 MiniCPM-V 4.0 | ModelBest / OpenBMB | 4.1B | 32,768 configured positions | Image + text | Apache 2.0 (model card) | Publisher source ↗ |
How to read the entries
A value is shown only for fields checked against the linked publisher source in this review. “Undisclosed” means the publisher source does not state the value. “Unverified” means this catalog has not checked the value against a primary source. Neither label implies that the model lacks the capability.
Evaluate for your task and device
- Describe one real task and define what a useful answer must contain. Include representative inputs, such as a note to summarize or a document image to read.
- Open the publisher source and confirm the exact model variant, license, input types, supported runtime, and any required conversion or quantization. Do not assume sibling variants share these properties.
- Install a supported build on the actual target device. Use the same representative inputs and task-specific prompts for each candidate, while following each model’s required chat format.
- Record task success, errors, elapsed time, memory use, and any battery or thermal effects you can measure. Repeat the tasks and keep the device, runtime, settings, and input set consistent.
A result from one phone, runtime, or quantized file does not establish compatibility or performance on another. This catalog does not claim that any listed model runs on phones.