Set up private local chat with AtomicChat’s compact LFM3 model on your phone. Use OnDevice on iPhone, or follow the documented Android import steps with exact-model compatibility clearly marked as unverified.
Choose a compatible iPhone running iOS 18 or later with sufficient available memory. Speed and model capacity depend on the device.
| Machine | AD-Q4_K_M · 1.13 GB |
|---|---|
| iPhone · OnDevice · iOS 18+ | Use the Q4 file with sufficient available memory. Reduce context or close other apps if memory is tight. |
| Android · Layla / ChatterUI | Documented local GGUF import; exact LFM3 checkpoint compatibility remains unverified. |
Allow at least 3 GB free as an estimated download/import allowance. Layla’s internal-storage import makes a copy. Storage is not a runtime-RAM requirement.
Install the compact model on your phone and use it for private conversations, short summaries, and rewriting. Choose your phone below.
Use the same 1.13 GB file throughout this guide. A larger quantization is not a workaround for missing LFM3 support.
LFM3-1.7B-Uncensored-AD-Q4_K_M.gguf
Plan for at least 3 GB free as an estimated download/import allowance, for the download and imported copy. The app, model, context and operating system also need working memory.
For iPhone, use OnDevice on iOS 18 or later with sufficient available memory. Android offers documented import procedures with exact-model compatibility still unverified.
Follow these steps on a compatible iPhone running iOS 18 or later with enough available memory. Start with Q4 and reduce context if memory is tight. Model loading and speed vary by device.
Use the linked App Store listing by MESUT CAN YAGCI and complete the initial setup. It requires iOS 18 or later. The current app listing documents LFM3 support. Choose phone-local inference and leave optional web or cloud features off for private local chat.
OnDevice on the App StoreIn OnDevice, open the “HuggingFace Search” area. Search for the repository name below, open the matching AtomicChat result, and choose LFM3-1.7B-Uncensored-AD-Q4_K_M.gguf (about 1.13 GB). Download it directly in the app and wait for the download to finish before continuing.
Select the downloaded model and open a new chat. If a context setting is exposed, start at 4,096 tokens as a conservative suggestion, then reduce it if the app runs out of memory. Keep the model’s embedded chat template where supported and begin with the app’s default sampling settings. Context and speed were not measured in the user’s test.
Confirm that the selected model can answer a new question without internet access.
After the model loads, send the prompt below. Then ask “Rewrite sentence three for a ten-year-old.” A useful result is a fresh response followed by a rewrite that refers to the earlier answer. Check the facts yourself; a small uncensored model can still be confidently wrong.
Keep the local model selected, disable optional online services, and turn off Wi-Fi and cellular data in the phone’s settings. Ask a new prompt rather than re-opening an old answer. A newly generated response demonstrates offline inference in this setup. Reconnect when you need downloads or updates; local conversations can continue without internet access.
Continue a conversation while its history fits your context setting, or start a new chat to clear it. Use the app’s stop-generation control to interrupt a reply and unload the model when finished if that control is available. Next time, open the app and select the downloaded model again; it should not need another download. Stored chat history follows the app’s own save/delete settings.
Send a new question with Wi-Fi and cellular off, then ask a follow-up that uses the previous answer. Stop a reply whenever you need to, start a fresh chat for a new topic, and reopen the downloaded model next time.
Sources checked October 30, 2026. The confirmed OnDevice run used an iPhone 16; the instructions apply to compatible iPhones generally. Android compatibility with this exact model remains unverified. “Uncensored” describes reduced refusals, not factual accuracy.
Check the exact filename and the app’s LFM3 architecture support. GGUF import alone is not sufficient. Update the app and confirm memory headroom. An unsupported-architecture error requires an updated runtime; downloading the larger Q6 file is not a fix.
Stop generation, let the phone cool, close other memory-heavy apps, and reduce context if the setting is exposed. Use the Q4 file and a short new conversation. The model download size is smaller than total runtime memory use.
Start a new conversation, check the model’s chat-template handling, and return sampling settings to the app defaults. Confirm the runtime supports this LFM3 build; changing a system prompt cannot repair missing architecture support.
Original instructions, model files, and compatibility notes behind this setup.
Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.
Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.
Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.