Edge AI inference platforms and on‑device model frameworks encompass the software and model ecosystems that enable vision and multimodal AI to run with low latency, constrained power, and limited connectivity. As of 2025‑12‑15, demand for on‑device inference is driven by privacy and data‑sovereignty requirements, real‑time autonomy use cases, specialized hardware (NPUs, embedded GPUs), and operational needs in defense and industrial contexts. Key trends include model efficiency (quantization, pruning, distilled/foundation models tuned for edge), sensor‑fusion LBMs for real‑time reasoning, deterministic middleware for safety‑critical systems, and hybrid orchestration across edge and cloud. Representative tools illustrate these patterns: Mistral AI provides open, efficiency‑focused foundation models and an enterprise production stack emphasizing governance and local deployment; Archetype AI’s Newton is positioned as a Large Behavior Model for multimodal sensor fusion and on‑prem/edge inference; Shield AI supplies autonomy software (Hivemind) plus EdgeOS middleware and tools (Pilot, Forge) for deterministic, mission‑critical autonomy; Run:ai (NVIDIA Run:ai) offers Kubernetes‑native GPU pooling and orchestration to maximize utilization across on‑prem, hybrid, and cloud edge clusters; no‑code platforms like Anakin.ai and Duckie lower the barrier to build and deploy domain‑specific agents and workflows using prebuilt apps or knowledge‑base agents. Together these capabilities reflect a practical stack for edge vision: efficient model formats and compilers, runtime frameworks for heterogeneous hardware, orchestration for distributed inference, and application tooling that shortens integration cycles — all anchored by governance, verification, and energy/latency tradeoffs essential for production deployments.