Drafts, edits, and summaries need different checks. Use the same brief and source material when you compare models.
We checked the linked vendor sources on October 14, 2026. This post reports no comparative product test.
Check local runtime and model compatibility
These are model candidates for a compatible local runtime. This post does not establish that each model fits an iPhone. Check the exact model file, available memory, and runtime support. These entries do not describe the PocketLLM catalog.
Qwen 2.5, Mistral 7B, Mistral Nemo, Llama 3.2, Phi 3.5 Mini and Gemma 2 for writing
1. Qwen2.5-7B-Instruct
Qwen publishes this instruction model under Apache 2.0. See the vendor model card. Ask for a short summary. Compare each claim with the source document.
2. Mistral-7B-Instruct-v0.3
Mistral publishes this instruction model under Apache 2.0. See the vendor model card. Ask for an email draft from a factual brief. Check that the draft adds no unsupported details.
3. Mistral-Nemo-Instruct-2407
Mistral publishes this instruction model under Apache 2.0. See the vendor model card. Ask for a document outline. Check the order and the required sections.
4. Llama-3.2-3B-Instruct
Meta publishes this text model under the Llama 3.2 Community License. See the vendor model card. Ask for a reply in a specified tone. Check the meaning before you send it.
5. Phi-3.5-mini-instruct
Microsoft publishes this instruction model under the MIT license. See the vendor model card. Ask for a set of steps. Check the sequence and all required conditions.
6. Gemma-2-9b-it
Google describes this model as an English text model under the Gemma terms. See the vendor model card. Ask for an English text edit. Check that the changes preserve the intended meaning.
PocketLLM is pre-launch. Private, on-device AI, starting on iPhone, with more platforms planned. No account required, no in-app tracking, and chat processing stays on your device. Join the waitlist.
Qwen, Mistral, Llama, Phi and Gemma: text checks
| Option | Proposed text check |
|---|---|
| 1. Qwen2.5-7B-Instruct | Check summary accuracy. |
| 2. Mistral-7B-Instruct-v0.3 | Check new factual claims. |
| 3. Mistral-Nemo-Instruct-2407 | Check document structure. |
| 4. Llama-3.2-3B-Instruct | Check tone and meaning. |
| 5. Phi-3.5-mini-instruct | Check procedural accuracy. |
| 6. Gemma-2-9b-it | Check the English edit. |
Before you choose
- Save the model version and configuration with your sample output.
- Check factual accuracy and compliance with your instructions.
- Compare the output with your original text.
- Select the model only after it passes your own requirements.
Local inference keeps prompts on the device during inference when the selected model runs there. For a local-only workflow, select no cloud model. Disable chat sync and cloud backup. Check the app telemetry policy. Disable optional network features. Review extensions, exports, and local storage access.
Frequently asked questions
Which license does Mistral-7B-Instruct-v0.3 use?
Mistral publishes this instruction model under Apache 2.0. See its model card.
Which license does Phi-3.5-mini-instruct use?
Microsoft publishes this instruction model under the MIT license. See its model card.
Does Google describe Gemma 2 9B as an English model?
Google describes Gemma 2 9B as an English text model under the Gemma terms. See the Google model card.
PocketLLM is pre-launch, with a waitlist for the iPhone launch. More platforms are planned. See the PocketLLM model catalog for its separate model list.