Running large language models on Android has long been a game of trial, error, and reading dense quantization tables. Off Grid AI Mobile (OGAM) is changing that dynamic with its latest update, introducing a curated selection system that automatically recommends models based on your device's specific RAM capacity. The feature, rolled out in version 0.0.116 and solidified in the current stable release 0.0.117, presents users with three distinct tiers: Fast, Balanced, and Best. This move effectively removes the cognitive load of model selection, allowing users to tap a card, download the file, and immediately switch their active chat model without ever needing to understand the difference between a Q4_K_M and a Q5_K_S file.

The Algorithm Behind the Tiers

OGAM’s selection logic is strictly bound by memory constraints rather than arbitrary benchmarks. The app calculates a safe operating budget by analyzing the phone's total RAM against the model's file size. The rule is precise: files under approximately 36% of total RAM are marked as running well, while those between 36% and 50% are flagged as 'Tight,' indicating potential instability or memory pressure. Anything exceeding 50% is simply not offered. For a standard 12 GB device, this means only models under 4.3 GB are considered safe for the 'Best' slot, with a hard cap around 6 GB. The 'Fast' tier specifically targets models that operate at the speed of a 4-billion parameter model or faster, often leveraging mixture-of-experts architectures where the active parameter count is low despite a larger total size.

Model Sources and Execution Details

The curated catalog draws from established model families including Qwen, Gemma, Llama, Mistral, and Phi. Crucially, the app filters out deprecated models and those incompatible with the user's current inference engine version. For example, a 12 GB phone might be served Qwen 3.5 variants in 0.8B, 2B, and 4B sizes. When running on the CPU or NPU, OGAM prioritizes Q4_0 quantized versions if available, as these formats offer superior performance on those specific backends. This technical nuance is hidden from the user; they simply see the resulting 'Fast' or 'Balanced' card. The system also extends to image generation and speech transcription, applying similar size-based logic to ensure that visual and audio models also fit within the device's memory envelope.

Limitations and Manual Control

While the automation is convenient, it is not a magic bullet. The picks are derived from static catalog metadata and RAM calculations, not real-time benchmarking on the specific hardware. OGAM explicitly states that it does not test speed on the user's device during the selection process. Users seeking granular control can still bypass these recommendations entirely. The full catalog remains accessible below the tiered cards, and a 'Sizes' section in model details allows users to manually select 'Smaller,' 'Recommended,' or 'Best' file variants. Additionally, the app supports manual refreshes to pull new models from the catalog without requiring a full app update, ensuring that the 'Best' slot can be upgraded as newer, more efficient models are released by the open-source community.

Key Takeaways

  • OGAM 0.0.117 introduces automated model selection based on RAM capacity, categorizing picks into Fast, Balanced, and Best tiers.
  • The algorithm restricts model downloads to files under 36% of total RAM for optimal performance, with a 'Tight' warning for files up to 50%.
  • Selections are sourced from major model families like Qwen and Llama, with automatic preference for Q4_0 quantization on CPU/NPU backends.
  • The system does not perform real-time speed benchmarks on the device, relying instead on static catalog data and file size metrics.

The Bottom Line

This is a necessary abstraction for the mobile AI market. Most users do not care about quantization methods; they care about latency and stability. By hiding the complexity behind a simple RAM check, OGAM makes local AI viable for the average Android user rather than just power users. Opinionated take: The era of manually configuring llama.cpp flags on Android is over for the mainstream. If you are still reading quantization tables to run a 7B model, you are doing it wrong.