With the rapid advancement of large language models (LLMs), we often encounter situations where new models like Qwen3 are not natively supported by existing hardware or platforms (e.g., Qwen3 not supported on NP8). In such cases, how can we adapt or enable these models on our platforms?
Are there recommended strategies such as model conversion, simplification, or the implementation of compatibility layers to achieve successful deployment?
What are the typical workflows or toolkits used to address these support gaps, and are fallback mechanisms (such as running on CPU or alternative accelerators) viable options when platform-native support is lacking?
Thank you for this inquiry.
MediaTek Genio AI (GAI) Toolkit applies model, weight, and operator conversions to optimize supported LLMs for deployment on the NPU (Neural Processing Unit). These steps are designed to maximize model performance on Genio platforms.
For models not yet officially supported—such as Qwen3 on NP8—the toolkit does not provide streamlined quick-adaptation routines, and CPU fallback or alternative accelerator deployments are not currently enabled for GAI models. MediaTek recommends following official support lists for model compatibility, as platform-native adaptation workflows presently require native toolkit integration and performance optimization.
Community members with special model requirements or suggestions are invited to submit feedback via Genio AI Forum Pinned Announcement. MediaTek regularly reviews input to inform future roadmap and documentation updates :)!