Choose a model
Find a comfortable fit for your phone.
Blobs runs a model on the phone to write replies. It searches enabled libraries and supplies relevant passages to that model. It contains no hosted inference client.
Available models
The bundled list includes Qwen3.5 2B, Phi-4 Mini, LFM2 1.2B, and Qwen2.5 0.5B for 2 GB-class phones. The table below reads the same metadata used by the Android app.
| Model | Download | Minimum RAM |
|---|---|---|
| Qwen3.5 2B Q4_K_M · 262k | 1.28 GB | 6 GiB |
| Phi-4 Mini Q4_K_M · 128k | 2.49 GB | 8 GiB |
| LFM2 1.2B Q4_K_M · 128k | 730.9 MB | 4 GiB |
| Qwen2.5 0.5B Q4_K_M · 32k | 491.4 MB | 2 GiB |
Minimum RAM is the list’s compatibility floor, not a guarantee of a particular speed. Android, other apps, prompt length, temperature, and available memory all affect performance. Installation needs room for the download, verification, and the installed model; follow the space guidance in the app.
Pick a comfortable starting point
The default is Qwen3.5 2B (Q4_K_M, about 1.28 GB), the official Unsloth GGUF with a declared 262k context window; non-thinking chat is used by default. Phi-4 Mini (Q4_K_M, about 2.49 GB) remains available as a larger alternative. LFM2 1.2B (Q4_K_M, about 0.73 GB, Liquid AI's official GGUF, 128k declared) is the compact alternative, and Qwen2.5 0.5B (Q4_K_M, about 0.46 GB) is the lightest option for low-memory phones. On the test emulator LFM2 ran a grammar-constrained turn at 50.3 against 92.3 milliseconds per token and held 1.30 against 1.89 GiB of resident memory, which is the figure that matters on a 4 GB phone. Those are emulator numbers from one device, kept as a comparison between models rather than a speed promise; the same turns have since run on a physical Android phone. Qwen3.5 2B remains the default on any phone that meets its declared memory floor, and below that floor the welcome recommendation resolves to the model that can actually load there. The Android runtime fixes the active native context at 4,096 tokens and the answer budget at 200 tokens; it does not select a larger context from device RAM. A model's advertised context is not allocated in full.
Change models
Open Settings → Models. Download another compatible option, then select the installed model you want to use. Downloading and switching are separate actions. Remove a model you no longer need to reclaim storage; keep another working model available for chat.
Tap Check for updates when connected to refresh the available choices. Opening the screen doesn’t start a background download.
Bringing a model file
Add a model file is available in model settings for local packages. Being a GGUF file does not by itself make a model compatible: architecture, tokenizer, prompt formatting, and memory requirements must also fit the pinned runtime. The in-app choices are the supported starting point.
Reducing local work
Smaller models, focused questions, and a narrower library scope reduce the amount of local work. The app has a Stop control for an active reply.
When the app moves to the background, it stops an active reply but keeps a loaded model for a warm return. Changing the selected model or explicitly trimming inference memory releases its weights.