• Run any model—text, vision, audio, speech, image-gen—fully on-device
• Works on NPU, GPU, or CPU with built-in support for Qualcomm & Apple NPUs, GGUF, and Apple MLX
• Supports SOTA models like Gemma-3n, PaddleOCR, Qwen3, Parakeet v3, Phi-4, and our own OmniNeural-4B
No fragmented runtimes, no complex setup—just one lightweight toolkit from prototype to production.
We were recently featured at Snapdragon Summit 2025 and are working with Qualcomm, AMD, and Intel to make on-device AI more accessible.
Would love feedback from the HN community—performance numbers, build experiences, feature requests, anything.
We’re also celebrating our Product Hunt launch today: https://www.producthunt.com/products/nexa-sdk-2