Nexa SDK

Nexa SDK is an on-device AI framework for developers building apps that run models on users' hardware. It covers language, vision and audio tasks through a shared interface for local inference.

• Multimodal conversations: Compatible models such as Qwen3-VL, Gemma-3n and OmniNeural-4B handle text, images or audio, with support for comparing multiple images or audio clips.
• Speech and image generation: Parakeet transcribes speech, including live microphone input, Kokoro generates speech, and SDXL creates images on supported hardware.
• Application integration: An OpenAI-compatible API server provides streamed responses and function calling defined with JSON schemas.
• Model choice: The SDK runs compatible GGUF, MLX and .nexa models from Hugging Face, and its command-line interface supports local chat and downloaded model management.

It runs on macOS, Windows and Linux using CPUs, GPUs or supported NPUs. GPU backends include CUDA, Metal and Vulkan; NPU support includes Qualcomm, Intel and AMD. GGUF works on all three desktop platforms, while MLX requires Apple Silicon macOS. Qualcomm NPU inference requires a Snapdragon X Elite laptop. Apple Neural Engine support covers Parakeet speech recognition.

Inference runs locally, but some models require a Nexa account and license token. Compatibility depends on the model format and hardware backend. Some MLX models from mlx-community may run unreliably; Nexa also provides a curated model collection.