IFA 2025 in Berlin is set to spotlight one of the most buzzed-about themes in AI: running large models directly on devices instead of relying on the cloud. OptAI, a South Korea–based startup, is stepping into the global spotlight with its Opt-Hancer platform, designed to supercharge on-device AI performance across CPUs, GPUs, and NPUs.
The company’s pitch is bold: “Any Model, Any Device.”
Cloud-Free AI: From Snapdragon to Multimodal Models
At IFA, OptAI will demonstrate how Opt-Hancer optimizes a wide spectrum of AI workloads:
- Large Language Models: LLaMA 3, Gemma 2, Qwen 2.5, EXAONE 4.0
- Vision Models: DeTR, YOLO family
- Multimodal Models: Qwen2-VL, LLaVA
One highlight: the optimization of LG EXAONE 3.5 for the Qualcomm Snapdragon 8 Elite NPU, built in collaboration with LG Uplus. That project is now gearing up for beta testing and commercialization in LG’s upcoming ixi-O service based on EXAONE 4.0.
OptAI is also showing results across the Snapdragon 8 Gen 2, Gen 3, and Elite chipsets, underscoring its ability to squeeze maximum performance from next-gen NPUs.
From Pitch Battle to DreamStage
OptAI will be busy in Berlin:
- Pitch Battle – live demos showcasing how Opt-Hancer optimizes models for diverse silicon, cutting costs and latency.
- DreamStage – a vision presentation on how on-device AI enables data security, energy efficiency, and customizability, freeing enterprises from heavy cloud dependence.
Why It Matters
As enterprises explore edge AI to reduce cloud costs, improve privacy, and enable real-time responsiveness, startups like OptAI are racing to build the tools that make it practical. Competitors like Qualcomm, Apple, and MediaTek are already embedding NPUs in consumer devices, but the missing piece has been cross-model, cross-hardware optimization. OptAI is positioning itself as that glue layer.
By leaning on South Korea’s manufacturing strength and aiming at global device makers, OptAI wants to do for on-device AI what NVIDIA CUDA did for GPUs: become the default standard.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
