F2 · Edge AI & Agents
Multimodal VLM with MNN on Armv9
Overview
On the Armv9 CPU of the Radxa Orion O6, this guide natively builds KleidiAI-optimized MNN, loads the pre-packaged Qwen2.5-Omni-7B-MNN model, and validates local text, image and audio inference. The workflow merges a shelf image and a spoken restock instruction into a single multimodal inference, producing a structured ticket with coverage, restock zones, quantities, deadlines and substitution policy. This CPU-first approach chains the portable MNN runtime, Armv9 optimization and a real edge application, giving developers a local multimodal baseline that needs no GPU/NPU.
Scope of verification
- It is a CPU-only reference workflow on a single Radxa Orion O6, using the pre-packaged Qwen2.5-Omni MNN model and local assets to validate multimodal inference and structured output.
- Hardware/software dependencies include Armv9 Linux, a native MNN/KleidiAI build, ~15 GB of model assets and at least 32 GB of free storage; actual output content and quality still depend on model, prompt and input assets.
- Scope centers on runtime build and application-level multimodal validation; it does not cover model export, format conversion, quantization, GPU/NPU heterogeneous scheduling, cross-platform benchmarking or production integration.
Canonical source: https://learn.arm.com/learning-paths/cross-platform/multimodel_mnn_v9/
This page is a guided summary with boundary notes, not an official reproduction.
This page is a guided summary with boundary notes, not an official reproduction.