Features of ailia LLM

In this page, we present the features that are provided by both the C and the C# APIs.

Loading GGUF

It is possible to load GGUF files and conduct chat.

Utilizing the GPU

GPU inference is possible using Metal on iOS and macOS, and Vulkan on Windows.

Image input (VLM)

Ask questions about images with a vision-capable model and its matching mmproj. Image files, encoded image buffers and raw RGB are supported. Check vision capability after opening the projector. See API Usage for a calling example.

Audio input

Generate text from audio using an audio-capable model and its matching mmproj. Input may be a WAV/MP3/FLAC file or encoded buffer. Check the audio capability after opening the projector; vision-only models do not support audio. Raw PCM and audio synthesis are unsupported. See API Usage for the procedure.

Android QNN uses precompiled linear graphs from an audio-enabled mmproj .qnn. The encoder still includes CPU processing; it does not run entirely on the NPU.

Android QNN

Load a precompiled SoC-specific text .qnn and its matching mmproj .qnn for VLM/ALM. Source GGUF files are not required on the device. On QNN-enabled builds, query the device model stem and select packages for the same SoC and model. See API Usage.

Tool Use (Function Calling)

For models whose chat template supports tool calling (e.g. Gemma 4), OpenAI-compatible tool definitions can be passed so that the model calls tools (functions). The tool call output is guaranteed by a grammar and the raw output can be converted into an OpenAI-compatible assistant message (content / reasoning_content / tool_calls). After generation, retrieve structured assistant JSON and append it to history; append tool results as messages with role "tool".